def find_bug_introducing_commits(self, bug_fixing_commits,
commits_to_ignore, tokenized):
if tokenized:
db_path = TOKENIZED_BUG_INTRODUCING_COMMITS_DB
repo_dir = self.tokenized_git_repo_dir
else:
db_path = BUG_INTRODUCING_COMMITS_DB
repo_dir = self.git_repo_dir
def git_to_mercurial(rev):
if tokenized:
return self.tokenized_git_to_mercurial[rev]
else:
return vcs_map.git_to_mercurial(rev)
def mercurial_to_git(rev):
if tokenized:
return self.mercurial_to_tokenized_git[rev]
else:
return vcs_map.mercurial_to_git(rev)
logger.info("Download previously found bug-introducing commits...")
if db.is_old_version(db_path) or not db.exists(db_path):
db.download(db_path, force=True)
logger.info("Get previously found bug-introducing commits...")
prev_bug_introducing_commits = list(db.read(db_path))
prev_bug_introducing_commits_nodes = set(
bug_introducing_commit["bug_fixing_rev"]
for bug_introducing_commit in prev_bug_introducing_commits)
logger.info(
f"Already classified {len(prev_bug_introducing_commits)} commits..."
)
hashes_to_ignore = set(commit["rev"] for commit in commits_to_ignore)
with open("git_hashes_to_ignore", "w") as f:
f.writelines("{}\n".format(mercurial_to_git(commit["rev"]))
for commit in commits_to_ignore if not tokenized
or commit["rev"] in self.mercurial_to_tokenized_git)
logger.info(f"{len(bug_fixing_commits)} commits to analyze")
# Skip already found bug-introducing commits.
bug_fixing_commits = [
bug_fixing_commit for bug_fixing_commit in bug_fixing_commits if
bug_fixing_commit["rev"] not in prev_bug_introducing_commits_nodes
]
logger.info(
f"{len(bug_fixing_commits)} commits left to analyze after skipping already analyzed ones"
)
bug_fixing_commits = [
bug_fixing_commit for bug_fixing_commit in bug_fixing_commits
if bug_fixing_commit["rev"] not in hashes_to_ignore
]
logger.info(
f"{len(bug_fixing_commits)} commits left to analyze after skipping the ones in the ignore list"
)
if tokenized:
bug_fixing_commits = [
bug_fixing_commit for bug_fixing_commit in bug_fixing_commits
if bug_fixing_commit["rev"] in self.mercurial_to_tokenized_git
]
logger.info(
f"{len(bug_fixing_commits)} commits left to analyze after skipping the ones with no git hash"
)
def _init(git_repo_dir):
thread_local.git = GitRepository(git_repo_dir)
def find_bic(bug_fixing_commit):
logger.info("Analyzing {}...".format(bug_fixing_commit["rev"]))
git_fix_revision = mercurial_to_git(bug_fixing_commit["rev"])
commit = thread_local.git.get_commit(git_fix_revision)
# Skip huge changes, we'll likely be wrong with them.
if len(commit.modifications) > MAX_MODIFICATION_NUMBER:
logger.info("Skipping {} as it is too big".format(
bug_fixing_commit["rev"]))
return None
bug_introducing_modifications = thread_local.git.get_commits_last_modified_lines(
commit,
hashes_to_ignore_path=os.path.realpath("git_hashes_to_ignore"))
logger.info("Found {} for {}".format(bug_introducing_modifications,
bug_fixing_commit["rev"]))
bug_introducing_commits = []
for bug_introducing_hashes in bug_introducing_modifications.values(
):
for bug_introducing_hash in bug_introducing_hashes:
try:
bug_introducing_commits.append({
"bug_fixing_rev":
bug_fixing_commit["rev"],
"bug_introducing_rev":
git_to_mercurial(bug_introducing_hash),
})
except Exception as e:
# Skip commits that are in git but not in mercurial, as they are too old (older than "Free the lizard").
if not str(e).startswith(
"Missing git commit in the VCS map"):
raise
# Add an empty result, just so that we don't reanalyze this again.
if len(bug_introducing_commits) == 0:
bug_introducing_commits.append({
"bug_fixing_rev":
bug_fixing_commit["rev"],
"bug_introducing_rev":
"",
})
return bug_introducing_commits
bug_fixing_commits_queue = bug_fixing_commits.copy()
with concurrent.futures.ThreadPoolExecutor(initializer=_init,
initargs=(repo_dir, ),
max_workers=os.cpu_count() +
1) as executor:
def results():
num_analyzed = 0
# Analyze up to 500 commits at a time, to avoid the task running out of time.
while len(
bug_fixing_commits_queue) != 0 and num_analyzed != 500:
bug_introducing_commit_futures = []
for _ in range(
min(500 - num_analyzed,
len(bug_fixing_commits_queue))):
bug_introducing_commit_futures.append(
executor.submit(find_bic,
bug_fixing_commits_queue.pop()))
logger.info(
f"Analyzing a chunk of {len(bug_introducing_commit_futures)} commits"
)
for future in tqdm(
concurrent.futures.as_completed(
bug_introducing_commit_futures),
total=len(bug_introducing_commit_futures),
):
result = future.result()
if result is not None:
num_analyzed += 1
yield from result
db.append(db_path, results())
zstd_compress(db_path)
return len(bug_fixing_commits_queue) == 0
def generate_test_scheduling_history(self, granularity):
push_data_path = f"push_data_{granularity}.json"
updated = download_check_etag(
test_scheduling.PUSH_DATA_URL.format(granularity=granularity))
if updated:
zstd_decompress(push_data_path)
assert os.path.exists(
push_data_path), "Decompressed push data file exists"
# Get the commits DB.
assert db.download(repository.COMMITS_DB)
HISTORY_DATE_START = datetime.now() - relativedelta(
months=TRAINING_MONTHS[granularity])
if granularity == "label":
test_scheduling_db = test_scheduling.TEST_LABEL_SCHEDULING_DB
past_failures_db = os.path.join(
"data", test_scheduling.PAST_FAILURES_LABEL_DB)
elif granularity == "group":
test_scheduling_db = test_scheduling.TEST_GROUP_SCHEDULING_DB
past_failures_db = os.path.join(
"data", test_scheduling.PAST_FAILURES_GROUP_DB)
touched_together_db = os.path.join(
"data", test_scheduling.TOUCHED_TOGETHER_DB)
db.download(test_scheduling_db, support_files_too=True)
last_node = None
for test_data in test_scheduling.get_test_scheduling_history(
granularity):
last_node = test_data["revs"][0]
def generate_all_data():
past_failures = test_scheduling.get_past_failures(granularity)
push_num = past_failures[
"push_num"] if "push_num" in past_failures else 0
# We can start once we get to the last revision we added in the previous run.
can_start = True if last_node is None else False
commit_map = {}
for commit_data in tqdm(repository.get_commits()):
if not can_start:
if last_node == commit_data["node"]:
can_start = True
continue
commit_map[commit_data["node"]] = commit_data
with open(push_data_path, "r") as f:
push_data = json.load(f)
logger.info(f"push data nodes: {len(push_data)}")
if granularity == "label":
push_data = [(
revisions,
rename_tasks(push_tasks),
rename_tasks(possible_regressions),
rename_tasks(likely_regressions),
) for revisions, push_tasks, possible_regressions,
likely_regressions in push_data]
# In the last 28 pushes, we definitely run all possible runnables.
all_runnables_set = set(
sum((push_runnables
for _, push_runnables, _, _ in push_data[-28:]), []))
# Filter runnables we don't need.
all_runnables = filter_runnables(list(all_runnables_set),
all_runnables_set, granularity)
all_runnables_set = set(all_runnables_set)
logger.info(
f"{len(all_runnables_set)} runnables run in the last 28 pushes"
)
# Store all runnables in the past_failures DB so it can be used in the evaluation phase.
past_failures["all_runnables"] = all_runnables
# XXX: Should we recreate the DB from scratch if the previous all_runnables are not the
# same as the current ones?
saved_nodes = set()
skipped_no_commits = 0
skipped_too_big_commits = 0
skipped_no_runnables = 0
# We can start once we get to the last revision we added in the previous run.
can_start = True if last_node is None else False
if granularity == "group":
update_touched_together_gen = test_scheduling.update_touched_together(
)
next(update_touched_together_gen)
for i in tqdm(range(len(push_data))):
(
revisions,
push_runnables,
possible_regressions,
likely_regressions,
) = push_data.pop(0)
if not can_start:
if last_node == revisions[0]:
can_start = True
continue
push_num += 1
# XXX: Some commits are skipped in the repository mining, e.g. merges and backouts. Maybe we should not skip them.
commits = tuple(
commit_map.pop(revision) for revision in revisions
if revision in commit_map)
if len(commits) == 0:
skipped_no_commits += 1
continue
merged_commits = commit_features.merge_commits(commits)
# XXX: For now, skip commits which are too large.
# In the future we can either:
# - Improve shelve perf and go back to consider all files;
# - Consider only files which appear with a given frequency, like the "files" feature in commit_features;
# - Keep a limit of number of files.
if len(merged_commits["files"]) > 50:
skipped_too_big_commits += 1
continue
# If we considered all_runnables, we'd generate a huge amount of data.
# So we consider only the runnables which run in this push, and the possible and likely regressions
# from this push.
runnables_to_consider = list(
set(push_runnables + possible_regressions +
likely_regressions))
runnables_to_consider = filter_runnables(
runnables_to_consider, all_runnables_set, granularity)
if len(runnables_to_consider) == 0:
skipped_no_runnables += 1
continue
# Sync DB every 250 pushes, so we cleanup the shelve cache (we'd run OOM otherwise!).
if i % 250 == 0:
past_failures.sync()
pushdate = dateutil.parser.parse(merged_commits["pushdate"])
if granularity == "group":
update_touched_together_gen.send(commits[0]["node"])
for data in test_scheduling.generate_data(
past_failures,
merged_commits,
push_num,
runnables_to_consider,
possible_regressions,
likely_regressions,
):
if pushdate > HISTORY_DATE_START:
saved_nodes.add(i)
data["revs"] = revisions
yield data
if granularity == "group":
try:
update_touched_together_gen.send(None)
except StopIteration:
pass
logger.info(f"saved push data nodes: {len(saved_nodes)}")
logger.info(f"skipped {skipped_no_commits} (no commits in our DB)")
logger.info(f"skipped {skipped_too_big_commits} (too big commits)")
logger.info(
f"skipped {skipped_no_runnables} (no interesting runnables)")
past_failures["push_num"] = push_num
past_failures.close()
db.append(test_scheduling_db, generate_all_data())
zstd_compress(test_scheduling_db)
with open_tar_zst(past_failures_db) as tar:
tar.add(past_failures_db[:-len(".tar.zst")])
if granularity == "group":
with open_tar_zst(touched_together_db) as tar:
tar.add(touched_together_db[:-len(".tar.zst")])
def generate_push_data(self, runnable):
# We keep in the cache the fact that we failed to analyze a push for 10
# days, so if we re-run often we don't retry the same pushes many times.
MISSING_CACHE_RETENTION = 10 * 24 * 60
# We'll use the past TRAINING_MONTHS months only for training the model,
# but we use half TRAINING_MONTHS months more than that to calculate the
# failure statistics.
from_months = TRAINING_MONTHS[runnable] + math.floor(
TRAINING_MONTHS[runnable] / 2)
pushes = mozci.push.make_push_objects(
from_date=f"today-{from_months}month",
to_date="today-3day",
branch="autoland",
)
start_time = time.monotonic()
num_cached = 0
push_data = []
def cache_key(push):
return f"push_data.{runnable}.{push.rev}"
# Regenerating a large amount of data when we update the mozci regression detection
# algorithm is currently pretty slow, so we only regenerate 1000 pushes whenever we
# run.
to_regenerate = set()
"""for push in pushes[::-1]:
cached = adr.config.cache.get(cache_key(push))
if not cached:
continue
value, mozci_version = cached
if mozci_version != MOZCI_VERSION and len(to_regenerate) < 1000:
to_regenerate.add(value[0][0])"""
for push in tqdm(pushes):
key = cache_key(push)
if adr.config.cache.has(key) and push.revs[0] not in to_regenerate:
num_cached += 1
cached = adr.config.cache.get(key)
if cached:
value, mozci_version = cached
push_data.append(value)
else:
logger.info(f"Analyzing {push.rev} at the {runnable} level...")
try:
if runnable == "label":
runnables = push.task_labels
elif runnable == "group":
runnables = push.group_summaries.keys()
value = [
push.revs,
list(runnables),
list(push.get_possible_regressions(runnable)),
list(push.get_likely_regressions(runnable)),
]
push_data.append(value)
adr.config.cache.forever(key, (value, MOZCI_VERSION))
except adr.errors.MissingDataError:
logger.warning(
f"Tasks for push {push.rev} can't be found on ActiveData"
)
adr.config.cache.put(key, (), MISSING_CACHE_RETENTION)
except Exception:
traceback.print_exc()
adr.config.cache.put(key, (), MISSING_CACHE_RETENTION)
if time.monotonic() - start_time >= 10800:
self.upload_adr_cache()
start_time = time.monotonic()
logger.info(
f"{num_cached} pushes were already cached out of {len(pushes)}")
with open(f"push_data_{runnable}.json", "w") as f:
json.dump(push_data, f)
zstd_compress(f"push_data_{runnable}.json")
def generate_push_data(self, granularity: str, training_months: int,
reretrieve: int) -> None:
# We'll use the past training_months months only for training the model,
# but we use half training_months months more than that to calculate the
# failure statistics.
from_months = training_months + math.floor(training_months / 2)
# We use the actual date instead of 'today-X' aliases to avoid adr caching
# this query.
from_date = datetime.utcnow() - relativedelta(months=from_months)
to_date = datetime.utcnow() - relativedelta(days=3)
pushes = mozci.push.make_push_objects(
from_date=from_date.strftime("%Y-%m-%d"),
to_date=to_date.strftime("%Y-%m-%d"),
branch="autoland",
)
if granularity == "label":
push_data_db = test_scheduling.PUSH_DATA_LABEL_DB
elif granularity == "group":
push_data_db = test_scheduling.PUSH_DATA_GROUP_DB
elif granularity == "config_group":
push_data_db = test_scheduling.PUSH_DATA_CONFIG_GROUP_DB
def cache_key(push: mozci.push.Push) -> str:
return f"push_data.{granularity}.{push.rev}"
def generate(
futures: List[concurrent.futures.Future],
) -> Generator[PushResult, None, None]:
nonlocal reretrieve
num_cached = 0
num_pushes = len(pushes)
for _ in tqdm(range(num_pushes)):
push = pushes.pop(0)
cached = futures.pop(0).result()
semaphore.release()
# Regenerating a large amount of data when we update the mozci regression detection
# algorithm is currently pretty slow, so we only regenerate a subset of pushes whenever we
# run.
if cached:
value, mozci_version = cached
# Regenerate results which were generated with an older version of mozci.
if reretrieve > 0 and mozci_version != MOZCI_VERSION:
cached = None
reretrieve -= 1
# Regenerate results which don't contain the fix revision.
elif len(value) != 5:
cached = None
if cached:
num_cached += 1
value, mozci_version = cached
assert len(value) == 5
yield value
else:
logger.info(
f"Analyzing {push.rev} at the {granularity} level...")
key = cache_key(push)
try:
if granularity == "label":
runnables = push.task_labels
elif granularity == "group":
runnables = push.group_summaries.keys()
elif granularity == "config_group":
runnables = push.config_group_summaries.keys()
value = (
tuple(push.revs),
push.backedoutby or push.bustage_fixed_by,
tuple(runnables),
tuple(push.get_possible_regressions(granularity)),
tuple(push.get_likely_regressions(granularity)),
)
adr.config.cache.put(
key,
(value, MOZCI_VERSION),
adr.config["cache"]["retention"],
)
assert len(value) == 5
yield value
except adr.errors.MissingDataError:
logger.warning(
f"Tasks for push {push.rev} can't be found on ActiveData"
)
except Exception:
traceback.print_exc()
logger.info(
f"{num_cached} pushes were already cached out of {num_pushes}")
semaphore = threading.BoundedSemaphore(256)
def retrieve_from_cache(push):
semaphore.acquire()
return adr.config.cache.get(cache_key(push))
with concurrent.futures.ThreadPoolExecutor() as executor:
futures = [
executor.submit(retrieve_from_cache, push) for push in pushes
]
try:
db.write(push_data_db, generate(futures))
except Exception:
for f in futures:
f.cancel()
try:
semaphore.release()
except ValueError:
continue
raise
zstd_compress(push_data_db)
def generate_push_data(self, runnable):
# We keep in the cache the fact that we failed to analyze a push for 10
# days, so if we re-run often we don't retry the same pushes many times.
MISSING_CACHE_RETENTION = 10 * 24 * 60
# We'll use the past TRAINING_MONTHS months only for training the model,
# but we use half TRAINING_MONTHS months more than that to calculate the
# failure statistics.
from_months = TRAINING_MONTHS[runnable] + math.floor(
TRAINING_MONTHS[runnable] / 2)
# We use the actual date instead of 'today-X' aliases to avoid adr caching
# this query.
from_date = datetime.utcnow() - relativedelta(months=from_months)
to_date = datetime.utcnow() - relativedelta(days=3)
pushes = mozci.push.make_push_objects(
from_date=from_date.strftime("%Y-%m-%d"),
to_date=to_date.strftime("%Y-%m-%d"),
branch="autoland",
)
num_cached = 0
push_data = []
def cache_key(push):
return f"push_data.{runnable}.{push.rev}"
# Regenerating a large amount of data when we update the mozci regression detection
# algorithm is currently pretty slow, so we only regenerate 1000 pushes whenever we
# run.
to_regenerate = set()
"""for push in pushes[::-1]:
cached = adr.config.cache.get(cache_key(push))
if not cached:
continue
value, mozci_version = cached
if mozci_version != MOZCI_VERSION and len(to_regenerate) < 1000:
to_regenerate.add(value[0][0])"""
def periodically_upload_adr_cache():
start_time = time.monotonic()
while not upload_thread_stop.isSet():
if time.monotonic() - start_time >= 10800:
self.upload_adr_cache()
start_time = time.monotonic()
upload_thread_stop.wait(timeout=7)
upload_thread = threading.Thread(target=periodically_upload_adr_cache)
upload_thread_stop = threading.Event()
upload_thread.start()
s3_store = adr.util.cache_stores.S3Store({
"bucket": "communitytc-bugbug",
"prefix": "data/adr_cache/",
})
s3_store.set_serializer(CompressedPickleSerializer())
for push in tqdm(pushes):
key = cache_key(push)
if adr.config.cache.has(key) and push.revs[0] not in to_regenerate:
num_cached += 1
cached = adr.config.cache.get(key)
if cached:
s3_store.put(key, cached, adr.config["cache"]["retention"])
value, mozci_version = cached
push_data.append(value)
else:
logger.info(f"Analyzing {push.rev} at the {runnable} level...")
try:
if runnable == "label":
runnables = push.task_labels
elif runnable == "group":
runnables = push.group_summaries.keys()
value = [
push.revs,
list(runnables),
list(push.get_possible_regressions(runnable)),
list(push.get_likely_regressions(runnable)),
]
push_data.append(value)
adr.config.cache.put(key, (value, MOZCI_VERSION),
adr.config["cache"]["retention"])
s3_store.put(key, (value, MOZCI_VERSION),
adr.config["cache"]["retention"])
except adr.errors.MissingDataError:
logger.warning(
f"Tasks for push {push.rev} can't be found on ActiveData"
)
adr.config.cache.put(key, (), MISSING_CACHE_RETENTION)
except Exception:
traceback.print_exc()
adr.config.cache.put(key, (), MISSING_CACHE_RETENTION)
upload_thread_stop.set()
upload_thread.join()
logger.info(
f"{num_cached} pushes were already cached out of {len(pushes)}")
with open(f"push_data_{runnable}.json", "w") as f:
json.dump(push_data, f)
zstd_compress(f"push_data_{runnable}.json")
def find_bug_introducing_commits(
self, bug_fixing_commits, commits_to_ignore, tokenized
):
if tokenized:
db_path = TOKENIZED_BUG_INTRODUCING_COMMITS_DB
repo_dir = self.tokenized_git_repo_dir
else:
db_path = BUG_INTRODUCING_COMMITS_DB
repo_dir = self.git_repo_dir
def git_to_mercurial(rev):
if tokenized:
return self.tokenized_git_to_mercurial[rev]
else:
return vcs_map.git_to_mercurial(rev)
def mercurial_to_git(rev):
if tokenized:
return self.mercurial_to_tokenized_git[rev]
else:
return vcs_map.mercurial_to_git(rev)
logger.info("Download previously found bug-introducing commits...")
if db.is_old_version(db_path) or not db.exists(db_path):
db.download(db_path, force=True)
logger.info("Get previously found bug-introducing commits...")
prev_bug_introducing_commits = list(db.read(db_path))
prev_bug_introducing_commits_nodes = set(
bug_introducing_commit["bug_fixing_rev"]
for bug_introducing_commit in prev_bug_introducing_commits
)
logger.info(
f"Already classified {len(prev_bug_introducing_commits)} commits..."
)
hashes_to_ignore = set(commit["rev"] for commit in commits_to_ignore)
with open("git_hashes_to_ignore", "w") as f:
f.writelines(
"{}\n".format(mercurial_to_git(commit["rev"]))
for commit in commits_to_ignore
if not tokenized or commit["rev"] in self.mercurial_to_tokenized_git
)
logger.info(f"{len(bug_fixing_commits)} commits to analyze")
# Skip already found bug-introducing commits.
bug_fixing_commits = [
bug_fixing_commit
for bug_fixing_commit in bug_fixing_commits
if bug_fixing_commit["rev"] not in prev_bug_introducing_commits_nodes
]
logger.info(
f"{len(bug_fixing_commits)} commits left to analyze after skipping already analyzed ones"
)
bug_fixing_commits = [
bug_fixing_commit
for bug_fixing_commit in bug_fixing_commits
if bug_fixing_commit["rev"] not in hashes_to_ignore
]
logger.info(
f"{len(bug_fixing_commits)} commits left to analyze after skipping the ones in the ignore list"
)
if tokenized:
bug_fixing_commits = [
bug_fixing_commit
for bug_fixing_commit in bug_fixing_commits
if bug_fixing_commit["rev"] in self.mercurial_to_tokenized_git
]
logger.info(
f"{len(bug_fixing_commits)} commits left to analyze after skipping the ones with no git hash"
)
# Analyze up to 500 commits at a time, to avoid the task running out of time.
done = True
if len(bug_fixing_commits) > 500:
bug_fixing_commits = bug_fixing_commits[-500:]
done = False
with open("done", "w") as f:
f.write(str(1 if done else 0))
def _init(git_repo_dir):
global GIT_REPO
GIT_REPO = GitRepository(git_repo_dir)
def find_bic(bug_fixing_commit):
logger.info("Analyzing {}...".format(bug_fixing_commit["rev"]))
git_fix_revision = mercurial_to_git(bug_fixing_commit["rev"])
commit = GIT_REPO.get_commit(git_fix_revision)
# Skip huge changes, we'll likely be wrong with them.
if len(commit.modifications) > MAX_MODIFICATION_NUMBER:
return [None]
bug_introducing_modifications = GIT_REPO.get_commits_last_modified_lines(
commit, hashes_to_ignore_path=os.path.realpath("git_hashes_to_ignore")
)
logger.info(
"Found {} for {}".format(
bug_introducing_modifications, bug_fixing_commit["rev"]
)
)
bug_introducing_commits = []
for bug_introducing_hashes in bug_introducing_modifications.values():
for bug_introducing_hash in bug_introducing_hashes:
try:
bug_introducing_commits.append(
{
"bug_fixing_rev": bug_fixing_commit["rev"],
"bug_introducing_rev": git_to_mercurial(
bug_introducing_hash
),
}
)
except Exception as e:
# Skip commits that are in git but not in mercurial, as they are too old (older than "Free the lizard").
if not str(e).startswith("Missing git commit in the VCS map"):
raise
# Add an empty result, just so that we don't reanalyze this again.
if len(bug_introducing_commits) == 0:
bug_introducing_commits.append(
{
"bug_fixing_rev": bug_fixing_commit["rev"],
"bug_introducing_rev": "",
}
)
return bug_introducing_commits
with concurrent.futures.ThreadPoolExecutor(
initializer=_init, initargs=(repo_dir,), max_workers=os.cpu_count() + 1
) as executor:
bug_introducing_commits = executor.map(find_bic, bug_fixing_commits)
bug_introducing_commits = tqdm(
bug_introducing_commits, total=len(bug_fixing_commits)
)
bug_introducing_commits = list(
itertools.chain.from_iterable(bug_introducing_commits)
)
total_results_num = len(bug_introducing_commits)
bug_introducing_commits = list(filter(None, bug_introducing_commits))
logger.info(
f"Skipped {total_results_num - len(bug_introducing_commits)} commits as they were too big"
)
db.append(db_path, bug_introducing_commits)
zstd_compress(db_path)
def compress_and_upload():
zstd_compress(db_path)
db.upload(db_path)
def go(self):
logger.info(
"Generate map of bug ID -> bug data for all bugs which were defects"
)
bug_fixing_commits = list(db.read(BUG_FIXING_COMMITS_DB))
bug_fixing_commits_nodes = set(
bug_fixing_commit["rev"]
for bug_fixing_commit in bug_fixing_commits
if bug_fixing_commit["type"] in ("d", "r")
)
all_bug_ids = set(
commit["bug_id"]
for commit in repository.get_commits()
if commit["node"] in bug_fixing_commits_nodes
)
bug_map = {}
for bug in bugzilla.get_bugs():
if bug["id"] not in all_bug_ids:
continue
bug_map[bug["id"]] = bug
logger.info(
"Generate a map from function to the three last bugs which were fixed by touching that function"
)
past_bugs_by_function = {}
for commit in tqdm(repository.get_commits()):
if commit["node"] not in bug_fixing_commits_nodes:
continue
if commit["bug_id"] not in bug_map:
continue
bug = bug_map[commit["bug_id"]]
bug_str = "Bug {} - {}".format(bug["id"], bug["summary"])
for path, f_group in commit["functions"].items():
if path not in past_bugs_by_function:
past_bugs_by_function[path] = {}
for f in f_group:
if f[0] not in past_bugs_by_function[path]:
bugs_deque = deque([bug_str], maxlen=3)
else:
bugs_deque = past_bugs_by_function[path][f[0]]["bugs"]
bugs_deque.append(bug_str)
past_bugs_by_function[path][f[0]] = {
"start": f[1],
"end": f[2],
"bugs": bugs_deque,
}
with open("data/past_bugs_by_function.pickle", "wb") as f:
pickle.dump(past_bugs_by_function, f, protocol=pickle.HIGHEST_PROTOCOL)
zstd_compress("data/past_bugs_by_function.pickle")
def generate_test_scheduling_history(self):
updated = download_check_etag(PUSH_DATA_URL)
if updated:
zstd_decompress("push_data.json")
assert os.path.exists(
"push_data.json"), "Decompressed push data file exists"
# Get the commits DB.
assert db.download(repository.COMMITS_DB)
HISTORY_DATE_START = datetime.now() - relativedelta(
months=TRAINING_MONTHS)
db.download(test_scheduling.TEST_SCHEDULING_DB, support_files_too=True)
last_node = None
for test_data in test_scheduling.get_test_scheduling_history():
last_node = test_data["revs"][0]
def generate_all_data():
past_failures = test_scheduling.get_past_failures()
push_num = past_failures[
"push_num"] if "push_num" in past_failures else 0
# We can start once we get to the last revision we added in the previous run.
can_start = True if last_node is None else False
commit_map = {}
for commit_data in tqdm(repository.get_commits()):
if not can_start:
if last_node == commit_data["node"]:
can_start = True
continue
commit_map[commit_data["node"]] = commit_data
with open("push_data.json", "r") as f:
push_data = json.load(f)[1:]
logger.info(f"push data nodes: {len(push_data)}")
# In the last 28 pushes, we definitely run all possible tasks.
all_tasks_set = set(
sum((push_tasks for _, push_tasks, _, _ in push_data[-28:]),
[]))
# Filter tasks we don't need.
all_tasks = filter_tasks(list(all_tasks_set), all_tasks_set)
all_tasks_set = set(all_tasks)
logger.info(
f"{len(all_tasks_set)} tasks run in the last 28 pushes")
# Store all tasks in the past_failures DB so it can be used in the evaluation phase.
past_failures["all_tasks"] = all_tasks
# XXX: Should we recreate the DB from scratch if the previous all_tasks are not the
# same as the current ones?
saved_nodes = set()
skipped_no_commits = 0
skipped_too_big_commits = 0
skipped_no_tasks = 0
# We can start once we get to the last revision we added in the previous run.
can_start = True if last_node is None else False
for i in tqdm(range(len(push_data))):
(
revisions,
push_tasks,
possible_regressions,
likely_regressions,
) = push_data.pop(0)
if not can_start:
if last_node == revisions[0]:
can_start = True
continue
push_num += 1
# XXX: Some commits are skipped in the repository mining, e.g. merges and backouts. Maybe we should not skip them.
commits = tuple(
commit_map.pop(revision) for revision in revisions
if revision in commit_map)
if len(commits) == 0:
skipped_no_commits += 1
continue
merged_commits = commit_features.merge_commits(commits)
# XXX: For now, skip commits which are too large.
# In the future we can either:
# - Improve shelve perf and go back to consider all files;
# - Consider only files which appear with a given frequency, like the "files" feature in commit_features;
# - Keep a limit of number of files.
if len(merged_commits["files"]) > 50:
skipped_too_big_commits += 1
continue
# If we considered all_tasks, we'd generate a huge amount of data.
# So we consider only the tasks which run in this push, and the possible and likely regressions
# from this push.
tasks_to_consider = list(
set(push_tasks + possible_regressions +
likely_regressions))
tasks_to_consider = filter_tasks(tasks_to_consider,
all_tasks_set)
if len(tasks_to_consider) == 0:
skipped_no_tasks += 1
continue
# Sync DB every 250 pushes, so we cleanup the shelve cache (we'd run OOM otherwise!).
if i % 250 == 0:
past_failures.sync()
pushdate = dateutil.parser.parse(merged_commits["pushdate"])
for data in test_scheduling.generate_data(
past_failures,
merged_commits,
push_num,
tasks_to_consider,
possible_regressions,
likely_regressions,
):
if pushdate > HISTORY_DATE_START:
saved_nodes.add(i)
data["revs"] = revisions
yield data
logger.info(f"saved push data nodes: {len(saved_nodes)}")
logger.info(f"skipped {skipped_no_commits} (no commits in our DB)")
logger.info(f"skipped {skipped_too_big_commits} (too big commits)")
logger.info(f"skipped {skipped_no_tasks} (no interesting tasks)")
past_failures["push_num"] = push_num
past_failures.close()
db.append(test_scheduling.TEST_SCHEDULING_DB, generate_all_data())
zstd_compress(test_scheduling.TEST_SCHEDULING_DB)
with open_tar_zst("data/past_failures.lmdb.tar.zst") as tar:
tar.add("data/past_failures.lmdb")
def generate_test_scheduling_history(self):
if not os.path.exists("push_data.json"):
download_check_etag(PUSH_DATA_URL, "push_data.json.zst")
zstd_decompress("push_data.json")
assert os.path.exists(
"push_data.json"
), "Decompressed push data file exists"
# Get the commits DB.
if db.is_old_version(repository.COMMITS_DB) or not db.exists(
repository.COMMITS_DB
):
db.download(repository.COMMITS_DB, force=True)
HISTORY_DATE_START = datetime.now() - relativedelta(months=TRAINING_MONTHS)
with open("push_data.json", "r") as f:
data = json.load(f)
push_data = {}
for row in data[1:]:
# Revision -> (all tasks, possible regressions, likely regressions)
push_data[row[0]] = (row[1], row[2], row[3])
logger.info(f"push data nodes: {len(push_data)}")
HISTORICAL_TIMESPAN = 56
if not db.is_old_version(test_scheduling.TEST_SCHEDULING_DB):
db.download(test_scheduling.TEST_SCHEDULING_DB, support_files_too=True)
for test_data in test_scheduling.get_test_scheduling_history():
pass
last_node = test_data["rev"]
else:
last_node = None
past_failures = shelve.open(
"data/past_failures.shelve",
protocol=pickle.HIGHEST_PROTOCOL,
writeback=True,
)
push_num = past_failures["push_num"] if "push_num" in past_failures else 0
def get_and_update_past_failures(type_, task, items, push_num, is_regression):
values_total = []
values_prev_7 = []
values_prev_14 = []
values_prev_28 = []
values_prev_56 = []
key = f"{type_}${task}$"
for item in items:
full_key = key + item
if full_key not in past_failures:
cur = past_failures[full_key] = ExpQueue(
push_num, HISTORICAL_TIMESPAN + 1, 0
)
else:
cur = past_failures[full_key]
value = cur[push_num]
values_total.append(value)
values_prev_7.append(value - cur[push_num - 7])
values_prev_14.append(value - cur[push_num - 14])
values_prev_28.append(value - cur[push_num - 28])
values_prev_56.append(value - cur[push_num - 56])
if is_regression:
cur[push_num] = value + 1
return (
sum(values_total),
sum(values_prev_7),
sum(values_prev_14),
sum(values_prev_28),
sum(values_prev_56),
)
def generate_data():
nonlocal push_num
commits_with_data = set()
saved_nodes = set()
# We can start once we get to the last revision we added in the previous run.
can_start = True if last_node is None else False
for commit_data in tqdm(repository.get_commits()):
node = commit_data["node"]
# Sync DB every 1000 commits, so we cleanup the shelve cache (we'd run OOM otherwise!).
if len(commits_with_data) % 1000 == 0:
past_failures.sync()
if node == last_node:
can_start = True
continue
if not can_start:
continue
if node not in push_data:
continue
commits_with_data.add(node)
commit_push_data = push_data[node]
for task in commit_push_data[0]:
if not any(task.startswith(j) for j in JOBS_TO_CONSIDER):
continue
is_regression = (
task in commit_push_data[1] or task in commit_push_data[2]
)
total_failures, past_7_pushes_failures, past_14_pushes_failures, past_28_pushes_failures, past_56_pushes_failures = get_and_update_past_failures(
"all", task, ["all"], push_num, is_regression
)
total_types_failures, past_7_pushes_types_failures, past_14_pushes_types_failures, past_28_pushes_types_failures, past_56_pushes_types_failures = get_and_update_past_failures(
"type", task, commit_data["types"], push_num, is_regression
)
total_files_failures, past_7_pushes_files_failures, past_14_pushes_files_failures, past_28_pushes_files_failures, past_56_pushes_files_failures = get_and_update_past_failures(
"file", task, commit_data["files"], push_num, is_regression
)
total_directories_failures, past_7_pushes_directories_failures, past_14_pushes_directories_failures, past_28_pushes_directories_failures, past_56_pushes_directories_failures = get_and_update_past_failures(
"directory",
task,
commit_data["directories"],
push_num,
is_regression,
)
total_components_failures, past_7_pushes_components_failures, past_14_pushes_components_failures, past_28_pushes_components_failures, past_56_pushes_components_failures = get_and_update_past_failures(
"component",
task,
commit_data["components"],
push_num,
is_regression,
)
pushdate = dateutil.parser.parse(commit_data["pushdate"])
if pushdate > HISTORY_DATE_START:
saved_nodes.add(node)
yield {
"rev": node,
"name": task,
"failures": total_failures,
"failures_past_7_pushes": past_7_pushes_failures,
"failures_past_14_pushes": past_14_pushes_failures,
"failures_past_28_pushes": past_28_pushes_failures,
"failures_past_56_pushes": past_56_pushes_failures,
"failures_in_types": total_types_failures,
"failures_past_7_pushes_in_types": past_7_pushes_types_failures,
"failures_past_14_pushes_in_types": past_14_pushes_types_failures,
"failures_past_28_pushes_in_types": past_28_pushes_types_failures,
"failures_past_56_pushes_in_types": past_56_pushes_types_failures,
"failures_in_files": total_files_failures,
"failures_past_7_pushes_in_files": past_7_pushes_files_failures,
"failures_past_14_pushes_in_files": past_14_pushes_files_failures,
"failures_past_28_pushes_in_files": past_28_pushes_files_failures,
"failures_past_56_pushes_in_files": past_56_pushes_files_failures,
"failures_in_directories": total_directories_failures,
"failures_past_7_pushes_in_directories": past_7_pushes_directories_failures,
"failures_past_14_pushes_in_directories": past_14_pushes_directories_failures,
"failures_past_28_pushes_in_directories": past_28_pushes_directories_failures,
"failures_past_56_pushes_in_directories": past_56_pushes_directories_failures,
"failures_in_components": total_components_failures,
"failures_past_7_pushes_in_components": past_7_pushes_components_failures,
"failures_past_14_pushes_in_components": past_14_pushes_components_failures,
"failures_past_28_pushes_in_components": past_28_pushes_components_failures,
"failures_past_56_pushes_in_components": past_56_pushes_components_failures,
"is_possible_regression": task in commit_push_data[1],
"is_likely_regression": task in commit_push_data[2],
}
# We no longer need the push data for this node, we can free the memory.
del push_data[node]
push_num += 1
logger.info(f"commits linked to push data: {len(commits_with_data)}")
logger.info(f"saved push data nodes: {len(saved_nodes)}")
db.append(test_scheduling.TEST_SCHEDULING_DB, generate_data())
zstd_compress(test_scheduling.TEST_SCHEDULING_DB)
past_failures["push_num"] = push_num
past_failures.close()
zstd_compress("data/past_failures.shelve")
def generate_push_data(self, pushes: List[mozci.push.Push],
granularity: str) -> None:
# We keep in the cache the fact that we failed to analyze a push for 10
# days, so if we re-run often we don't retry the same pushes many times.
MISSING_CACHE_RETENTION = 10 * 24 * 60
from_date = get_from_date(granularity)
pushes = [
push for push in pushes
if datetime.utcfromtimestamp(push.date) >= from_date
]
if granularity == "label":
push_data_db = test_scheduling.PUSH_DATA_LABEL_DB
elif granularity == "group":
push_data_db = test_scheduling.PUSH_DATA_GROUP_DB
elif granularity == "config_group":
push_data_db = test_scheduling.PUSH_DATA_CONFIG_GROUP_DB
def cache_key(push: mozci.push.Push) -> str:
return f"push_data.{granularity}.{push.rev}"
def generate(executor) -> Generator[PushResult, None, None]:
num_cached = 0
num_pushes = len(pushes)
# Regenerating a large amount of data when we update the mozci regression detection
# algorithm is currently pretty slow, so we only regenerate 1000 pushes whenever we
# run.
to_regenerate = 1000
semaphore = threading.BoundedSemaphore(256)
def retrieve_from_cache(push):
semaphore.acquire()
return adr.config.cache.get(cache_key(push))
futures = tuple(
executor.submit(retrieve_from_cache, push) for push in pushes)
for push, future in zip(tqdm(pushes), futures):
exc = future.exception()
if exc is not None:
logger.info(f"Exception {exc} while getting {push.rev}")
for f in futures:
f.cancel()
cached = future.result()
semaphore.release()
if cached and to_regenerate > 0:
value, mozci_version = cached
# Regenerate results which were generated when we were not cleaning
# up WPT groups.
if any(runnable.startswith("/") for runnable in value[1]):
cached = None
to_regenerate -= 1
"""# Regenerate results which were generated with an older version of mozci.
elif mozci_version != MOZCI_VERSION and to_regenerate > 0:
cached = None
to_regenerate -= 1"""
if cached is not None:
num_cached += 1
if cached:
value, mozci_version = cached
yield value
else:
logger.info(
f"Analyzing {push.rev} at the {granularity} level...")
key = cache_key(push)
try:
if granularity == "label":
runnables = push.task_labels
elif granularity == "group":
runnables = push.group_summaries.keys()
elif granularity == "config_group":
runnables = push.config_group_summaries.keys()
value = (
push.revs,
tuple(runnables),
tuple(push.get_possible_regressions(granularity)),
tuple(push.get_likely_regressions(granularity)),
)
adr.config.cache.put(
key,
(value, MOZCI_VERSION),
adr.config["cache"]["retention"],
)
yield value
except adr.errors.MissingDataError:
logger.warning(
f"Tasks for push {push.rev} can't be found on ActiveData"
)
adr.config.cache.put(key, (), MISSING_CACHE_RETENTION)
except Exception:
traceback.print_exc()
adr.config.cache.put(key, (), MISSING_CACHE_RETENTION)
logger.info(
f"{num_cached} pushes were already cached out of {num_pushes}")
with concurrent.futures.ThreadPoolExecutor() as executor:
db.write(push_data_db, generate(executor))
zstd_compress(push_data_db)
def generate_test_scheduling_history(self):
if not os.path.exists("push_data.json"):
download_check_etag(PUSH_DATA_URL, "push_data.json.zst")
zstd_decompress("push_data.json")
assert os.path.exists(
"push_data.json"), "Decompressed push data file exists"
# Get the commits DB.
if db.is_old_version(
repository.COMMITS_DB) or not db.exists(repository.COMMITS_DB):
db.download(repository.COMMITS_DB, force=True)
HISTORY_DATE_START = datetime.now() - relativedelta(
months=TRAINING_MONTHS)
HISTORICAL_TIMESPAN = 56
if not db.is_old_version(test_scheduling.TEST_SCHEDULING_DB):
db.download(test_scheduling.TEST_SCHEDULING_DB,
support_files_too=True)
for test_data in test_scheduling.get_test_scheduling_history():
pass
last_node = test_data["revs"][0]
else:
last_node = None
past_failures = shelve.Shelf(
LMDBDict("data/past_failures.lmdb"),
protocol=pickle.HIGHEST_PROTOCOL,
writeback=True,
)
push_num = past_failures[
"push_num"] if "push_num" in past_failures else 0
def get_and_update_past_failures(type_, task, items, push_num,
is_regression):
values_total = []
values_prev_7 = []
values_prev_14 = []
values_prev_28 = []
values_prev_56 = []
key = f"{type_}${task}$"
for item in items:
full_key = key + item
if full_key not in past_failures:
cur = past_failures[full_key] = ExpQueue(
push_num, HISTORICAL_TIMESPAN + 1, 0)
else:
cur = past_failures[full_key]
value = cur[push_num]
values_total.append(value)
values_prev_7.append(value - cur[push_num - 7])
values_prev_14.append(value - cur[push_num - 14])
values_prev_28.append(value - cur[push_num - 28])
values_prev_56.append(value - cur[push_num - 56])
if is_regression:
cur[push_num] = value + 1
return (
sum(values_total),
sum(values_prev_7),
sum(values_prev_14),
sum(values_prev_28),
sum(values_prev_56),
)
def generate_data():
nonlocal push_num
saved_nodes = set()
skipped_no_commits = 0
skipped_too_big_commits = 0
skipped_no_tasks = 0
# We can start once we get to the last revision we added in the previous run.
can_start = True if last_node is None else False
commit_map = {}
for commit_data in tqdm(repository.get_commits()):
if not can_start:
if last_node == commit_data["node"]:
can_start = True
continue
commit_map[commit_data["node"]] = commit_data
with open("push_data.json", "r") as f:
push_data = json.load(f)[1:]
logger.info(f"push data nodes: {len(push_data)}")
# In the last 28 pushes, we definitely run all possible tasks.
all_tasks_set = set(
sum((push_tasks for _, push_tasks, _, _ in push_data[-28:]),
[]))
# Filter tasks we don't need.
all_tasks = filter_tasks(list(all_tasks_set), all_tasks_set)
all_tasks_set = set(all_tasks)
logger.info(
f"{len(all_tasks_set)} tasks run in the last 28 pushes")
# We can start once we get to the last revision we added in the previous run.
can_start = True if last_node is None else False
for i in tqdm(range(len(push_data))):
(
revisions,
push_tasks,
possible_regressions,
likely_regressions,
) = push_data.pop(0)
if not can_start:
if last_node == revisions[0]:
can_start = True
continue
push_num += 1
# XXX: Some commits are skipped in the repository mining, e.g. merges and backouts. Maybe we should not skip them.
commits = tuple(
commit_map.pop(revision) for revision in revisions
if revision in commit_map)
if len(commits) == 0:
skipped_no_commits += 1
continue
merged_commits = commit_features.merge_commits(commits)
# XXX: For now, skip commits which are too large.
# In the future we can either:
# - Improve shelve perf and go back to consider all files;
# - Consider only files which appear with a given frequency, like the "files" feature in commit_features;
# - Keep a limit of number of files.
if len(merged_commits["files"]) > 20:
skipped_too_big_commits += 1
continue
# If we considered all_tasks, we'd generate a huge amount of data.
# So we consider only the tasks which run in this push, and the possible and likely regressions
# from this push.
tasks_to_consider = list(
set(push_tasks + possible_regressions +
likely_regressions))
tasks_to_consider = filter_tasks(tasks_to_consider,
all_tasks_set)
if len(tasks_to_consider) == 0:
skipped_no_tasks += 1
continue
# Sync DB every 250 pushes, so we cleanup the shelve cache (we'd run OOM otherwise!).
if i % 250 == 0:
past_failures.sync()
pushdate = dateutil.parser.parse(merged_commits["pushdate"])
for task in tasks_to_consider:
is_regression = (task in possible_regressions
or task in likely_regressions)
(
total_failures,
past_7_pushes_failures,
past_14_pushes_failures,
past_28_pushes_failures,
past_56_pushes_failures,
) = get_and_update_past_failures("all", task, ["all"],
push_num, is_regression)
(
total_types_failures,
past_7_pushes_types_failures,
past_14_pushes_types_failures,
past_28_pushes_types_failures,
past_56_pushes_types_failures,
) = get_and_update_past_failures("type", task,
merged_commits["types"],
push_num, is_regression)
(
total_files_failures,
past_7_pushes_files_failures,
past_14_pushes_files_failures,
past_28_pushes_files_failures,
past_56_pushes_files_failures,
) = get_and_update_past_failures("file", task,
merged_commits["files"],
push_num, is_regression)
(
total_directories_failures,
past_7_pushes_directories_failures,
past_14_pushes_directories_failures,
past_28_pushes_directories_failures,
past_56_pushes_directories_failures,
) = get_and_update_past_failures(
"directory",
task,
merged_commits["directories"],
push_num,
is_regression,
)
(
total_components_failures,
past_7_pushes_components_failures,
past_14_pushes_components_failures,
past_28_pushes_components_failures,
past_56_pushes_components_failures,
) = get_and_update_past_failures(
"component",
task,
merged_commits["components"],
push_num,
is_regression,
)
if pushdate > HISTORY_DATE_START:
saved_nodes.add(i)
yield {
"revs": revisions,
"name": task,
"failures": total_failures,
"failures_past_7_pushes": past_7_pushes_failures,
"failures_past_14_pushes": past_14_pushes_failures,
"failures_past_28_pushes": past_28_pushes_failures,
"failures_past_56_pushes": past_56_pushes_failures,
"failures_in_types": total_types_failures,
"failures_past_7_pushes_in_types":
past_7_pushes_types_failures,
"failures_past_14_pushes_in_types":
past_14_pushes_types_failures,
"failures_past_28_pushes_in_types":
past_28_pushes_types_failures,
"failures_past_56_pushes_in_types":
past_56_pushes_types_failures,
"failures_in_files": total_files_failures,
"failures_past_7_pushes_in_files":
past_7_pushes_files_failures,
"failures_past_14_pushes_in_files":
past_14_pushes_files_failures,
"failures_past_28_pushes_in_files":
past_28_pushes_files_failures,
"failures_past_56_pushes_in_files":
past_56_pushes_files_failures,
"failures_in_directories":
total_directories_failures,
"failures_past_7_pushes_in_directories":
past_7_pushes_directories_failures,
"failures_past_14_pushes_in_directories":
past_14_pushes_directories_failures,
"failures_past_28_pushes_in_directories":
past_28_pushes_directories_failures,
"failures_past_56_pushes_in_directories":
past_56_pushes_directories_failures,
"failures_in_components":
total_components_failures,
"failures_past_7_pushes_in_components":
past_7_pushes_components_failures,
"failures_past_14_pushes_in_components":
past_14_pushes_components_failures,
"failures_past_28_pushes_in_components":
past_28_pushes_components_failures,
"failures_past_56_pushes_in_components":
past_56_pushes_components_failures,
"is_possible_regression": task
in possible_regressions,
"is_likely_regression": task in likely_regressions,
}
logger.info(f"saved push data nodes: {len(saved_nodes)}")
logger.info(f"skipped {skipped_no_commits} (no commits in our DB)")
logger.info(f"skipped {skipped_too_big_commits} (too big commits)")
logger.info(f"skipped {skipped_no_tasks} (no interesting tasks)")
db.append(test_scheduling.TEST_SCHEDULING_DB, generate_data())
zstd_compress(test_scheduling.TEST_SCHEDULING_DB)
past_failures["push_num"] = push_num
past_failures.close()
with open_tar_zst("data/past_failures.lmdb.tar.zst") as tar:
tar.add("data/past_failures.lmdb")
def generate_push_data(self, granularity: str) -> None:
# We keep in the cache the fact that we failed to analyze a push for 10
# days, so if we re-run often we don't retry the same pushes many times.
MISSING_CACHE_RETENTION = 10 * 24 * 60
# We'll use the past TRAINING_MONTHS months only for training the model,
# but we use half TRAINING_MONTHS months more than that to calculate the
# failure statistics.
from_months = TRAINING_MONTHS[granularity] + math.floor(
TRAINING_MONTHS[granularity] / 2
)
# We use the actual date instead of 'today-X' aliases to avoid adr caching
# this query.
from_date = datetime.utcnow() - relativedelta(months=from_months)
to_date = datetime.utcnow() - relativedelta(days=3)
pushes = mozci.push.make_push_objects(
from_date=from_date.strftime("%Y-%m-%d"),
to_date=to_date.strftime("%Y-%m-%d"),
branch="autoland",
)
if granularity == "label":
push_data_db = test_scheduling.PUSH_DATA_LABEL_DB
elif granularity == "group":
push_data_db = test_scheduling.PUSH_DATA_GROUP_DB
cache: Dict[mozci.push.Push, Tuple[PushResult, int]] = {}
def cache_key(push: mozci.push.Push) -> str:
return f"push_data.{granularity}.{push.rev}"
with concurrent.futures.ThreadPoolExecutor() as executor:
future_to_push = {
executor.submit(
lambda push: adr.config.cache.get(cache_key(push)), push
): push
for push in pushes
}
for future in tqdm(
concurrent.futures.as_completed(future_to_push),
total=len(future_to_push),
):
push = future_to_push[future]
exc = future.exception()
if exc is not None:
logger.info(f"Exception {exc} while getting {push.rev}")
for f in future_to_push.keys():
f.cancel()
cache[push] = future.result()
# Regenerating a large amount of data when we update the mozci regression detection
# algorithm is currently pretty slow, so we only regenerate 1000 pushes whenever we
# run.
"""to_regenerate = 0
for push in pushes[::-1]:
cached = cache[push]
if not cached:
continue
value, mozci_version = cached
if mozci_version != MOZCI_VERSION and to_regenerate < 1000:
cache[push] = None
to_regenerate += 1"""
to_regenerate = 0
for push in pushes[::-1]:
cached = cache[push]
if not cached:
continue
if to_regenerate < 1000:
del cache[push]
adr.config.cache.put(push.push_uuid, {}, 0)
to_regenerate += 1
def generate() -> Generator[PushResult, None, None]:
num_cached = 0
for push in tqdm(pushes):
key = cache_key(push)
if push in cache and cache[push] is not None:
num_cached += 1
cached = cache[push]
if cached:
value, mozci_version = cached
yield value
else:
logger.info(f"Analyzing {push.rev} at the {granularity} level...")
try:
if granularity == "label":
runnables = push.task_labels
elif granularity == "group":
runnables = push.group_summaries.keys()
value = (
push.revs,
list(runnables),
list(push.get_possible_regressions(granularity)),
list(push.get_likely_regressions(granularity)),
)
adr.config.cache.put(
key,
(value, MOZCI_VERSION),
adr.config["cache"]["retention"],
)
yield value
except adr.errors.MissingDataError:
logger.warning(
f"Tasks for push {push.rev} can't be found on ActiveData"
)
adr.config.cache.put(key, (), MISSING_CACHE_RETENTION)
except Exception:
traceback.print_exc()
adr.config.cache.put(key, (), MISSING_CACHE_RETENTION)
logger.info(f"{num_cached} pushes were already cached out of {len(pushes)}")
db.write(push_data_db, generate())
zstd_compress(push_data_db)
def get_commits_to_ignore(self):
logger.info("Download previous commits to ignore...")
db.download(IGNORED_COMMITS_DB)
logger.info("Get previously classified commits...")
prev_commits_to_ignore = list(db.read(IGNORED_COMMITS_DB))
logger.info(f"Already found {len(prev_commits_to_ignore)} commits to ignore...")
# When we already have some analyzed commits, re-analyze the last 3500 to make sure
# we didn't miss back-outs that happened since the last analysis.
if len(prev_commits_to_ignore) > 0:
first_commit_to_reanalyze = (
-3500 if len(prev_commits_to_ignore) >= 3500 else 0
)
rev_start = "children({})".format(
prev_commits_to_ignore[first_commit_to_reanalyze]["rev"]
)
else:
rev_start = 0
with hglib.open(self.mercurial_repo_dir) as hg:
revs = repository.get_revs(hg, rev_start)
# Drop commits which are not yet present in the mercurial <-> git mapping.
while len(revs) > 0:
try:
vcs_map.mercurial_to_git(revs[-1].decode("ascii"))
break
except Exception as e:
if not str(e).startswith("Missing mercurial commit in the VCS map"):
raise
revs.pop()
commits = repository.hg_log_multi(self.mercurial_repo_dir, revs)
repository.set_commits_to_ignore(self.mercurial_repo_dir, commits)
chosen_commits = set()
commits_to_ignore = []
for commit in commits:
if commit.ignored or commit.backedoutby:
commits_to_ignore.append(
{
"rev": commit.node,
"type": "backedout" if commit.backedoutby else "",
}
)
chosen_commits.add(commit.node)
logger.info(f"{len(commits_to_ignore)} new commits to ignore...")
for prev_commit in prev_commits_to_ignore[::-1]:
if prev_commit["rev"] not in chosen_commits:
commits_to_ignore.append(prev_commit)
chosen_commits.add(prev_commit["rev"])
logger.info(f"{len(commits_to_ignore)} commits to ignore...")
logger.info(
"...of which {} are backed-out".format(
sum(1 for commit in commits_to_ignore if commit["type"] == "backedout")
)
)
db.write(IGNORED_COMMITS_DB, commits_to_ignore)
zstd_compress(IGNORED_COMMITS_DB)
db.upload(IGNORED_COMMITS_DB)
def retrieve_issues(
self, owner: str, repo: str, state: str, retrieve_events: bool
) -> None:
db.download(github.GITHUB_ISSUES_DB)
github.download_issues(owner, repo, state, retrieve_events)
zstd_compress(github.GITHUB_ISSUES_DB)
def generate_push_data(self, granularity: str, training_months: int,
reretrieve: int) -> None:
# We'll use the past training_months months only for training the model,
# but we use half training_months months more than that to calculate the
# failure statistics.
from_months = training_months + math.floor(training_months / 2)
# We use the actual date instead of 'today-X' aliases to avoid mozci caching
# this query.
from_date = datetime.utcnow() - relativedelta(months=from_months)
to_date = datetime.utcnow() - relativedelta(days=3)
if granularity == "label":
push_data_db = test_scheduling.PUSH_DATA_LABEL_DB
elif granularity == "group":
push_data_db = test_scheduling.PUSH_DATA_GROUP_DB
elif granularity == "config_group":
push_data_db = test_scheduling.PUSH_DATA_CONFIG_GROUP_DB
def cache_key(push: mozci.push.Push) -> str:
return f"push_data.{granularity}.{push.rev}"
def generate(
progress_bar: tqdm,
pushes: list[mozci.push.Push],
futures: list[concurrent.futures.Future],
) -> Generator[PushResult, None, None]:
nonlocal reretrieve
num_cached = 0
num_pushes = len(pushes)
num_errors = 0
for push, future in zip(pushes, futures):
cached = future.result()
# Regenerating a large amount of data when we update the mozci regression detection
# algorithm is currently pretty slow, so we only regenerate a subset of pushes whenever we
# run.
if cached:
value, mozci_version = cached
# Regenerate results which were generated with an older version of mozci.
if reretrieve > 0 and mozci_version != MOZCI_VERSION:
cached = None
reretrieve -= 1
if cached:
num_cached += 1
value, mozci_version = cached
assert len(value) == 5
if value != "ERROR":
yield value
else:
num_errors += 1
else:
logger.info(
f"Analyzing {push.rev} at the {granularity} level...")
key = cache_key(push)
try:
if granularity == "label":
runnables = push.label_summaries.keys()
elif granularity == "group":
runnables = push.group_summaries.keys()
elif granularity == "config_group":
runnables = push.config_group_summaries.keys()
value = (
tuple(push.revs),
push.backedoutby or push.bustage_fixed_by,
tuple(runnables),
tuple(push.get_possible_regressions(granularity)),
tuple(push.get_likely_regressions(granularity)),
)
mozci.config.cache.put(
key,
(value, MOZCI_VERSION),
mozci.config["cache"]["retention"],
)
assert len(value) == 5
yield value
except mozci.errors.MissingDataError:
logger.warning(
f"Tasks for push {push.rev} can't be found on ActiveData"
)
except Exception:
num_errors += 1
traceback.print_exc()
mozci.config.cache.put(
key,
("ERROR", MOZCI_VERSION),
mozci.config["cache"]["retention"],
)
progress_bar.update(1)
logger.info(
f"{num_cached} pushes were already cached out of {num_pushes}")
logger.info(f"There were errors in {num_errors} pushes")
def retrieve_from_cache(push):
return mozci.config.cache.get(cache_key(push))
total_pushes = len(
mozci.push.make_push_objects(
from_date=from_date.strftime("%Y-%m-%d"),
to_date=to_date.strftime("%Y-%m-%d"),
branch="autoland",
))
with concurrent.futures.ThreadPoolExecutor() as executor:
with tqdm(total=total_pushes) as progress_bar:
# Run in batches of 7 days to avoid running out of memory (given that mozci pushes
# consume a lot of memory, and they all have references to each other through "parent"
# and "child" links so they are basically never released while we run this).
while from_date < to_date:
next_from_date = from_date + relativedelta(days=7)
if next_from_date > to_date:
next_from_date = to_date
pushes = mozci.push.make_push_objects(
from_date=from_date.strftime("%Y-%m-%d"),
to_date=next_from_date.strftime("%Y-%m-%d"),
branch="autoland",
)
futures = [
executor.submit(retrieve_from_cache, push)
for push in pushes
]
try:
db.append(push_data_db,
generate(progress_bar, pushes, futures))
except Exception:
for f in futures:
f.cancel()
raise
from_date = next_from_date
zstd_compress(push_data_db)
def retrieve_bugs(self, limit=None):
bugzilla.set_token(get_secret("BUGZILLA_TOKEN"))
db.download(bugzilla.BUGS_DB)
# Get IDs of bugs changed since last run.
last_modified = db.last_modified(bugzilla.BUGS_DB)
logger.info(
f"Retrieving IDs of bugs modified since the last run on {last_modified}"
)
changed_ids = bugzilla.get_ids({
"f1": "delta_ts",
"o1": "greaterthaneq",
"v1": last_modified.date()
})
logger.info(f"Retrieved {len(changed_ids)} IDs.")
# Get IDs of bugs between (two years and six months ago) and (six months ago).
six_months_ago = datetime.utcnow() - relativedelta(months=6)
two_years_and_six_months_ago = six_months_ago - relativedelta(years=2)
logger.info(
f"Retrieving bug IDs from {two_years_and_six_months_ago} to {six_months_ago}"
)
timespan_ids = bugzilla.get_ids_between(two_years_and_six_months_ago,
six_months_ago)
if limit:
timespan_ids = timespan_ids[:limit]
logger.info(f"Retrieved {len(timespan_ids)} IDs.")
# Get IDs of labelled bugs.
labelled_bug_ids = labels.get_all_bug_ids()
if limit:
labelled_bug_ids = labelled_bug_ids[:limit]
logger.info(f"{len(labelled_bug_ids)} labelled bugs to download.")
# Get the commits DB, as we need it to get the bug IDs linked to recent commits.
# XXX: Temporarily avoid downloading the commits DB when a limit is set, to avoid the integration test fail when the commits DB is bumped.
if limit is None:
assert db.download(repository.COMMITS_DB)
# Get IDs of bugs linked to commits (used for some commit-based models, e.g. backout and regressor).
start_date = datetime.now() - relativedelta(years=2, months=6)
commit_bug_ids = [
commit["bug_id"] for commit in repository.get_commits()
if commit["bug_id"]
and dateutil.parser.parse(commit["pushdate"]) >= start_date
]
if limit:
commit_bug_ids = commit_bug_ids[-limit:]
logger.info(
f"{len(commit_bug_ids)} bugs linked to commits to download.")
# Get IDs of bugs which caused regressions fixed by commits (useful for the regressor model).
regressed_by_bug_ids = sum(
[
bug["regressed_by"]
for bug in bugzilla.get_bugs() if bug["id"] in commit_bug_ids
],
[],
)
if limit:
regressed_by_bug_ids = regressed_by_bug_ids[-limit:]
logger.info(
f"{len(regressed_by_bug_ids)} bugs which caused regressions fixed by commits."
)
all_ids = (timespan_ids + labelled_bug_ids + commit_bug_ids +
regressed_by_bug_ids)
all_ids_set = set(all_ids)
# We have to redownload bugs that were changed since the last download.
# We can remove from the DB the bugs that are outside of the considered timespan and are not labelled.
bugzilla.delete_bugs(lambda bug: bug["id"] in changed_ids or bug["id"]
not in all_ids_set)
bugzilla.download_bugs(all_ids)
# Get regressed_by_bug_ids again (the set could have changed after downloading new bugs).
regressed_by_bug_ids = sum(
[
bug["regressed_by"]
for bug in bugzilla.get_bugs() if bug["id"] in commit_bug_ids
],
[],
)
logger.info(
f"{len(regressed_by_bug_ids)} bugs which caused regressions fixed by commits."
)
bugzilla.download_bugs(regressed_by_bug_ids)
# Try to re-download inconsistent bugs, up to three times.
inconsistent_bugs = bugzilla.get_bugs(include_invalid=True)
for i in range(3):
# We look for inconsistencies in all bugs first, then, on following passes,
# we only look for inconsistencies in bugs that were found to be inconsistent in the first pass
inconsistent_bugs = bug_snapshot.get_inconsistencies(
inconsistent_bugs)
inconsistent_bug_ids = set(bug["id"] for bug in inconsistent_bugs)
if len(inconsistent_bug_ids) == 0:
break
logger.info(
f"Re-downloading {len(inconsistent_bug_ids)} bugs, as they were inconsistent"
)
bugzilla.delete_bugs(lambda bug: bug["id"] in inconsistent_bug_ids)
bugzilla.download_bugs(inconsistent_bug_ids)
zstd_compress("data/bugs.json")
def go(self) -> None:
logger.info(
"Generate map of bug ID -> bug data for all bugs which were defects"
)
bug_fixing_commits = list(db.read(BUG_FIXING_COMMITS_DB))
bug_fixing_commits_nodes = set(
bug_fixing_commit["rev"]
for bug_fixing_commit in bug_fixing_commits
if bug_fixing_commit["type"] in ("d", "r"))
logger.info(
f"{len(bug_fixing_commits_nodes)} bug-fixing commits to analyze")
all_bug_ids = set(commit["bug_id"]
for commit in repository.get_commits())
bug_map = {
bug["id"]: bug
for bug in bugzilla.get_bugs() if bug["id"] in all_bug_ids
}
logger.info(
"Generate a map from files/functions to the bugs which were fixed/introduced by touching them"
)
# TODO: Support "moving" past bugs between files when they are renamed and between functions when they are
# moved across files.
by_dimensions = ["file", "directory", "component"]
def dimension_to_field(dimension: str) -> str:
return f"{dimension}s" if dimension != "directory" else "directories"
past_regressions_by: dict[str, dict[str, list[int]]] = defaultdict(
lambda: defaultdict(list))
past_fixed_bugs_by: dict[str, dict[str, list[int]]] = defaultdict(
lambda: defaultdict(list))
past_regression_blocked_bugs_by: dict[str, dict[
str, list[int]]] = defaultdict(lambda: defaultdict(list))
past_fixed_bug_blocked_bugs_by: dict[str, dict[
str, list[int]]] = defaultdict(lambda: defaultdict(list))
past_regressions_by_function: dict[str, dict[
str, list[int]]] = defaultdict(lambda: defaultdict(list))
past_fixed_bugs_by_function: dict[str, dict[
str, list[int]]] = defaultdict(lambda: defaultdict(list))
past_regression_blocked_bugs_by_function: dict[str, dict[
str, list[int]]] = defaultdict(lambda: defaultdict(list))
past_fixed_bug_blocked_bugs_by_function: dict[str, dict[
str, list[int]]] = defaultdict(lambda: defaultdict(list))
for commit in tqdm(repository.get_commits()):
if commit["bug_id"] not in bug_map:
continue
if commit["backedoutby"]:
continue
bug = bug_map[commit["bug_id"]]
if len(bug["regressions"]) > 0:
for dimension in by_dimensions:
for path in commit[dimension_to_field(dimension)]:
past_regressions_by[dimension][path].extend(
bug_id for bug_id in bug["regressions"]
if bug_id in bug_map)
past_regression_blocked_bugs_by[dimension][
path].extend(bugzilla.find_blocked_by(
bug_map, bug))
for path, f_group in commit["functions"].items():
for f in f_group:
past_regressions_by_function[path][f["name"]].extend(
bug_id for bug_id in bug["regressions"]
if bug_id in bug_map)
past_regression_blocked_bugs_by_function[path][
f["name"]].extend(
bugzilla.find_blocked_by(bug_map, bug))
if commit["node"] in bug_fixing_commits_nodes:
for dimension in by_dimensions:
for path in commit[dimension_to_field(dimension)]:
past_fixed_bugs_by[dimension][path].append(bug["id"])
past_fixed_bug_blocked_bugs_by[dimension][path].extend(
bugzilla.find_blocked_by(bug_map, bug))
for path, f_group in commit["functions"].items():
for f in f_group:
past_fixed_bugs_by_function[path][f["name"]].append(
bug["id"])
past_fixed_bug_blocked_bugs_by_function[path][
f["name"]].extend(
bugzilla.find_blocked_by(bug_map, bug))
def _transform(bug_ids: list[int]) -> list[dict]:
seen = set()
results = []
for bug_id in bug_ids:
if bug_id in seen:
continue
seen.add(bug_id)
bug = bug_map[bug_id]
results.append({
"id":
bug_id,
"summary":
bug["summary"],
"component":
"{}::{}".format(bug["product"], bug["component"]),
})
return results
def past_bug_ids_to_summaries(
past_bugs_by: dict[str, list[int]]) -> dict[str, list[dict]]:
return {
path: _transform(bug_ids)
for path, bug_ids in past_bugs_by.items()
}
for dimension in by_dimensions:
with open(f"data/past_regressions_by_{dimension}.json", "w") as f:
json.dump(
past_bug_ids_to_summaries(past_regressions_by[dimension]),
f)
zstd_compress(f"data/past_regressions_by_{dimension}.json")
with open(f"data/past_fixed_bugs_by_{dimension}.json", "w") as f:
json.dump(
past_bug_ids_to_summaries(past_fixed_bugs_by[dimension]),
f)
zstd_compress(f"data/past_fixed_bugs_by_{dimension}.json")
with open(f"data/past_regression_blocked_bugs_by_{dimension}.json",
"w") as f:
json.dump(
past_bug_ids_to_summaries(
past_regression_blocked_bugs_by[dimension]),
f,
)
zstd_compress(
f"data/past_regression_blocked_bugs_by_{dimension}.json")
with open(f"data/past_fixed_bug_blocked_bugs_by_{dimension}.json",
"w") as f:
json.dump(
past_bug_ids_to_summaries(
past_fixed_bug_blocked_bugs_by[dimension]),
f,
)
zstd_compress(
f"data/past_fixed_bug_blocked_bugs_by_{dimension}.json")
def past_function_bug_ids_to_summaries(
past_bugs: dict[str, dict[str, list[int]]]
) -> dict[str, dict[str, list[dict]]]:
return {
path: {
func: _transform(bug_ids)
for func, bug_ids in funcs_bugs.items()
}
for path, funcs_bugs in past_bugs.items()
}
with open("data/past_regressions_by_function.json", "w") as f:
json.dump(
past_function_bug_ids_to_summaries(
past_regressions_by_function), f)
zstd_compress("data/past_regressions_by_function.json")
with open("data/past_fixed_bugs_by_function.json", "w") as f:
json.dump(
past_function_bug_ids_to_summaries(
past_fixed_bugs_by_function), f)
zstd_compress("data/past_fixed_bugs_by_function.json")
with open("data/past_regression_blocked_bugs_by_function.json",
"w") as f:
json.dump(
past_function_bug_ids_to_summaries(
past_regression_blocked_bugs_by_function),
f,
)
zstd_compress("data/past_regression_blocked_bugs_by_function.json")
with open("data/past_fixed_bug_blocked_bugs_by_function.json",
"w") as f:
json.dump(
past_function_bug_ids_to_summaries(
past_fixed_bug_blocked_bugs_by_function),
f,
)
zstd_compress("data/past_fixed_bug_blocked_bugs_by_function.json")
def find_bug_fixing_commits(self):
logger.info("Downloading commits database...")
assert db.download(repository.COMMITS_DB)
logger.info("Downloading bugs database...")
assert db.download(bugzilla.BUGS_DB)
logger.info("Download previous classifications...")
db.download(BUG_FIXING_COMMITS_DB)
logger.info("Get previously classified commits...")
prev_bug_fixing_commits = list(db.read(BUG_FIXING_COMMITS_DB))
prev_bug_fixing_commits_nodes = set(
bug_fixing_commit["rev"]
for bug_fixing_commit in prev_bug_fixing_commits)
logger.info(
f"Already classified {len(prev_bug_fixing_commits)} commits...")
# TODO: Switch to the pure Defect model, as it's better in this case.
logger.info("Downloading defect/enhancement/task model...")
defect_model = download_and_load_model("defectenhancementtask")
logger.info("Downloading regression model...")
regression_model = download_and_load_model("regression")
start_date = datetime.now() - RELATIVE_START_DATE
end_date = datetime.now() - RELATIVE_END_DATE
logger.info(
f"Gathering bug IDs associated to commits (since {start_date} and up to {end_date})..."
)
commit_map = defaultdict(list)
for commit in repository.get_commits():
if commit["node"] in prev_bug_fixing_commits_nodes:
continue
commit_date = dateutil.parser.parse(commit["pushdate"])
if commit_date < start_date or commit_date > end_date:
continue
commit_map[commit["bug_id"]].append(commit["node"])
logger.info(
f"{sum(len(commit_list) for commit_list in commit_map.values())} commits found, {len(commit_map)} bugs linked to commits"
)
assert len(commit_map) > 0
def get_relevant_bugs():
return (bug for bug in bugzilla.get_bugs()
if bug["id"] in commit_map)
bug_count = sum(1 for bug in get_relevant_bugs())
logger.info(
f"{bug_count} bugs in total, {len(commit_map) - bug_count} bugs linked to commits missing"
)
known_defect_labels = defect_model.get_labels()
known_regression_labels = regression_model.get_labels()
bug_fixing_commits = []
def append_bug_fixing_commits(bug_id, type_):
for commit in commit_map[bug_id]:
bug_fixing_commits.append({"rev": commit, "type": type_})
for bug in tqdm(get_relevant_bugs(), total=bug_count):
# Ignore bugs which are not linked to the commits we care about.
if bug["id"] not in commit_map:
continue
# If we know the label already, we don't need to apply the model.
if (bug["id"] in known_regression_labels
and known_regression_labels[bug["id"]] == 1):
append_bug_fixing_commits(bug["id"], "r")
continue
if bug["id"] in known_defect_labels:
if known_defect_labels[bug["id"]] == "defect":
append_bug_fixing_commits(bug["id"], "d")
else:
append_bug_fixing_commits(bug["id"], "e")
continue
if defect_model.classify(bug)[0] == "defect":
if regression_model.classify(bug)[0] == 1:
append_bug_fixing_commits(bug["id"], "r")
else:
append_bug_fixing_commits(bug["id"], "d")
else:
append_bug_fixing_commits(bug["id"], "e")
db.append(BUG_FIXING_COMMITS_DB, bug_fixing_commits)
zstd_compress(BUG_FIXING_COMMITS_DB)
db.upload(BUG_FIXING_COMMITS_DB)
def generate_push_data(self, runnable):
def upload_adr_cache():
cache_path = os.path.splitext(ADR_CACHE_DB)[0]
assert os.path.abspath(
adr.config["cache"]["stores"]["file"]["path"]
) == os.path.abspath(cache_path)
with open_tar_zst(f"{ADR_CACHE_DB}.zst") as tar:
tar.add(cache_path)
db.upload(ADR_CACHE_DB)
# We'll use the past TRAINING_MONTHS months only for training the model,
# but we use half TRAINING_MONTHS months more than that to calculate the
# failure statistics.
from_months = TRAINING_MONTHS[runnable] + math.floor(
TRAINING_MONTHS[runnable] / 2
)
pushes = mozci.push.make_push_objects(
from_date=f"today-{from_months}month",
to_date="today-3day",
branch="autoland",
)
start_time = time.monotonic()
num_cached = 0
push_data = []
for push in tqdm(pushes):
key = f"push_data.{runnable}.{push.rev}"
logger.info(f"Analyzing {push.rev} at the {runnable} level...")
if adr.config.cache.has(key):
num_cached += 1
push_data.append(adr.config.cache.get(key))
else:
try:
if runnable == "label":
runnables = push.task_labels
elif runnable == "group":
runnables = push.group_summaries.keys()
value = [
push.revs,
list(runnables),
list(push.get_possible_regressions(runnable)),
list(push.get_likely_regressions(runnable)),
]
push_data.append(value)
adr.config.cache.forever(key, value)
except adr.errors.MissingDataError:
logger.warning(
f"Tasks for push {push.rev} can't be found on ActiveData"
)
except Exception:
traceback.print_exc()
if time.monotonic() - start_time >= 3600:
upload_adr_cache()
start_time = time.monotonic()
logger.info(f"{num_cached} pushes were already cached out of {len(pushes)}")
upload_adr_cache()
with open(f"push_data_{runnable}.json", "w") as f:
json.dump(push_data, f)
zstd_compress(f"push_data_{runnable}.json")
def generate_test_scheduling_history(self, granularity: str,
training_months: int) -> None:
if granularity != "config_group":
# Get the commits DB.
assert db.download(repository.COMMITS_DB)
HISTORY_DATE_START = datetime.now() - relativedelta(
months=training_months)
if granularity == "label":
test_scheduling_db = test_scheduling.TEST_LABEL_SCHEDULING_DB
past_failures_db = os.path.join(
"data", test_scheduling.PAST_FAILURES_LABEL_DB)
failing_together_db = os.path.join(
"data", test_scheduling.FAILING_TOGETHER_LABEL_DB)
elif granularity == "group":
test_scheduling_db = test_scheduling.TEST_GROUP_SCHEDULING_DB
past_failures_db = os.path.join(
"data", test_scheduling.PAST_FAILURES_GROUP_DB)
touched_together_db = os.path.join(
"data", test_scheduling.TOUCHED_TOGETHER_DB)
elif granularity == "config_group":
test_scheduling_db = test_scheduling.TEST_CONFIG_GROUP_SCHEDULING_DB
past_failures_db = os.path.join(
"data", test_scheduling.PAST_FAILURES_CONFIG_GROUP_DB)
failing_together_db = os.path.join(
"data", test_scheduling.FAILING_TOGETHER_CONFIG_GROUP_DB)
push_data_iter, push_data_count, all_runnables = test_scheduling.get_push_data(
granularity)
if granularity in ("label", "config_group"):
test_scheduling.generate_failing_together_probabilities(
granularity, push_data_iter(), push_data_count)
def generate_all_data() -> Generator[Dict[str, Any], None, None]:
past_failures = test_scheduling.get_past_failures(
granularity, False)
push_num = past_failures[
"push_num"] if "push_num" in past_failures else 0
commit_map = {}
for commit_data in tqdm(repository.get_commits()):
commit_map[commit_data["node"]] = commit_data
# Store all runnables in the past_failures DB so it can be used in the evaluation phase.
past_failures["all_runnables"] = all_runnables
# XXX: Should we recreate the DB from scratch if the previous all_runnables are not the
# same as the current ones?
saved_nodes = set()
skipped_no_commits = 0
skipped_too_big_commits = 0
skipped_no_runnables = 0
if granularity in ("group", "config_group"):
update_touched_together_gen = test_scheduling.update_touched_together(
)
next(update_touched_together_gen)
for (
i,
(
revisions,
fix_revision,
push_runnables,
possible_regressions,
likely_regressions,
),
) in enumerate(tqdm(push_data_iter(), total=push_data_count)):
push_num += 1
# XXX: Some commits are skipped in the repository mining, e.g. merges and backouts. Maybe we should not skip them.
commits = tuple(
commit_map.pop(revision) for revision in revisions
if revision in commit_map)
if len(commits) == 0:
skipped_no_commits += 1
continue
merged_commits = commit_features.merge_commits(commits)
# XXX: For now, skip commits which are too large.
# In the future we can either:
# - Improve shelve perf and go back to consider all files;
# - Consider only files which appear with a given frequency, like the "files" feature in commit_features;
# - Keep a limit of number of files.
if len(merged_commits["files"]) > 50:
skipped_too_big_commits += 1
continue
# If we considered all_runnables, we'd generate a huge amount of data.
# We consider only the runnables which run in this push, and the possible and likely regressions
# from this push. We can't consider all runnables because we can't be sure that a task that didn't
# run on a push would have been successful.
runnables_to_consider = list(
set(push_runnables + possible_regressions +
likely_regressions))
if len(runnables_to_consider) == 0:
skipped_no_runnables += 1
continue
# Sync DB every 250 pushes, so we cleanup the shelve cache (we'd run OOM otherwise!).
if i % 250 == 0:
past_failures.sync()
pushdate = dateutil.parser.parse(merged_commits["pushdate"])
if granularity in ("group", "config_group"):
update_touched_together_gen.send(commits[0]["node"])
result_data = []
for data in test_scheduling.generate_data(
granularity,
past_failures,
merged_commits,
push_num,
runnables_to_consider,
possible_regressions,
likely_regressions,
):
if pushdate > HISTORY_DATE_START:
result_data.append(data)
if pushdate > HISTORY_DATE_START:
saved_nodes.add(i)
yield {
"revs": revisions,
"data": result_data,
}
if granularity == "group":
try:
update_touched_together_gen.send(None)
except StopIteration:
pass
logger.info(f"saved push data nodes: {len(saved_nodes)}")
logger.info(f"skipped {skipped_no_commits} (no commits in our DB)")
logger.info(f"skipped {skipped_too_big_commits} (too big commits)")
logger.info(
f"skipped {skipped_no_runnables} (no interesting runnables)")
past_failures["push_num"] = push_num
past_failures.close()
# For the config/group granularity, we are only interested in the failing together DB.
if granularity != "config_group":
db.append(test_scheduling_db, generate_all_data())
zstd_compress(test_scheduling_db)
create_tar_zst(past_failures_db)
if granularity == "group":
create_tar_zst(touched_together_db)
if granularity in ("label", "config_group"):
create_tar_zst(failing_together_db)
def compress_and_upload() -> None:
utils.zstd_compress(SHADOW_SCHEDULER_STATS_DB)
db.upload(SHADOW_SCHEDULER_STATS_DB)
def retrieve_bugs(self, limit: int = None) -> None:
bugzilla.set_token(get_secret("BUGZILLA_TOKEN"))
db.download(bugzilla.BUGS_DB)
# Get IDs of bugs changed since last run.
last_modified = db.last_modified(bugzilla.BUGS_DB)
logger.info(
f"Retrieving IDs of bugs modified since the last run on {last_modified}"
)
changed_ids = set(
bugzilla.get_ids({
"f1": "delta_ts",
"o1": "greaterthaneq",
"v1": last_modified.date()
}))
logger.info(f"Retrieved {len(changed_ids)} IDs.")
all_components = bugzilla.get_product_component_count(9999)
deleted_component_ids = set(
bug["id"] for bug in bugzilla.get_bugs() if "{}::{}".format(
bug["product"], bug["component"]) not in all_components)
logger.info(
f"{len(deleted_component_ids)} bugs belonging to deleted components"
)
changed_ids |= deleted_component_ids
# Get IDs of bugs between (two years and six months ago) and now.
two_years_and_six_months_ago = datetime.utcnow() - relativedelta(
years=2, months=6)
logger.info(f"Retrieving bug IDs since {two_years_and_six_months_ago}")
timespan_ids = bugzilla.get_ids_between(two_years_and_six_months_ago)
if limit:
timespan_ids = timespan_ids[-limit:]
logger.info(f"Retrieved {len(timespan_ids)} IDs.")
# Get IDs of labelled bugs.
labelled_bug_ids = labels.get_all_bug_ids()
if limit:
labelled_bug_ids = labelled_bug_ids[-limit:]
logger.info(f"{len(labelled_bug_ids)} labelled bugs to download.")
# Get the commits DB, as we need it to get the bug IDs linked to recent commits.
# XXX: Temporarily avoid downloading the commits DB when a limit is set, to avoid the integration test fail when the commits DB is bumped.
if limit is None:
assert db.download(repository.COMMITS_DB)
# Get IDs of bugs linked to commits (used for some commit-based models, e.g. backout and regressor).
start_date = datetime.now() - relativedelta(years=3)
commit_bug_ids = list(
set(commit["bug_id"] for commit in repository.get_commits()
if commit["bug_id"]
and dateutil.parser.parse(commit["pushdate"]) >= start_date))
if limit:
commit_bug_ids = commit_bug_ids[-limit:]
logger.info(
f"{len(commit_bug_ids)} bugs linked to commits to download.")
# Get IDs of bugs which are regressions, bugs which caused regressions (useful for the regressor model),
# and blocked bugs.
regression_related_ids: List[int] = list(
set(
sum(
(bug["regressed_by"] + bug["regressions"] + bug["blocks"]
for bug in bugzilla.get_bugs()),
[],
)))
if limit:
regression_related_ids = regression_related_ids[-limit:]
logger.info(
f"{len(regression_related_ids)} bugs which caused regressions fixed by commits."
)
# Get IDs of bugs linked to intermittent failures.
test_failure_bug_ids = [
item["bug_id"] for item in test_scheduling.get_failure_bugs(
two_years_and_six_months_ago, datetime.utcnow())
]
if limit:
test_failure_bug_ids = test_failure_bug_ids[-limit:]
logger.info(f"{len(test_failure_bug_ids)} bugs about test failures.")
all_ids = (timespan_ids + labelled_bug_ids + commit_bug_ids +
regression_related_ids + test_failure_bug_ids)
all_ids_set = set(all_ids)
# We have to redownload bugs that were changed since the last download.
# We can remove from the DB the bugs that are outside of the considered timespan and are not labelled.
bugzilla.delete_bugs(lambda bug: bug["id"] in changed_ids or bug["id"]
not in all_ids_set)
new_bugs = bugzilla.download_bugs(all_ids)
# Get regression_related_ids again (the set could have changed after downloading new bugs).
for i in range(7):
regression_related_ids = list(
set(
sum(
(bug["regressed_by"] + bug["regressions"] +
bug["blocks"] for bug in new_bugs),
[],
)))
logger.info(
f"{len(regression_related_ids)} bugs which caused regressions fixed by commits."
)
if limit:
regression_related_ids = regression_related_ids[-limit:]
# If we got all bugs we needed, break.
if set(regression_related_ids).issubset(all_ids):
break
new_bugs = bugzilla.download_bugs(regression_related_ids)
# Try to re-download inconsistent bugs, up to twice.
inconsistent_bugs = bugzilla.get_bugs(include_invalid=True)
for i in range(2):
# We look for inconsistencies in all bugs first, then, on following passes,
# we only look for inconsistencies in bugs that were found to be inconsistent in the first pass
inconsistent_bugs = bug_snapshot.get_inconsistencies(
inconsistent_bugs)
inconsistent_bug_ids = set(bug["id"] for bug in inconsistent_bugs)
if len(inconsistent_bug_ids) == 0:
break
logger.info(
f"Re-downloading {len(inconsistent_bug_ids)} bugs, as they were inconsistent"
)
bugzilla.delete_bugs(lambda bug: bug["id"] in inconsistent_bug_ids)
bugzilla.download_bugs(inconsistent_bug_ids)
# TODO: Figure out why.
missing_history_bug_ids = {
bug["id"]
for bug in bugzilla.get_bugs() if "history" not in bug
}
bugzilla.delete_bugs(lambda bug: bug["id"] in missing_history_bug_ids)
logger.info(
f"Deleted {len(missing_history_bug_ids)} bugs as we couldn't retrieve their history"
)
zstd_compress(bugzilla.BUGS_DB)
def retrieve_test_scheduling_history(self):
os.makedirs("data", exist_ok=True)
# Download previous cache.
cache_path = os.path.abspath("data/adr_cache")
if not os.path.exists(cache_path):
try:
download_check_etag(URL, "adr_cache.tar.xz")
with tarfile.open("adr_cache.tar.xz", "r:xz") as tar:
tar.extractall()
assert os.path.exists(
"data/adr_cache"), "Decompressed adr cache exists"
except requests.exceptions.HTTPError:
logger.info("The adr cache is not available yet")
# Setup adr cache configuration.
os.makedirs(os.path.expanduser("~/.config/adr"), exist_ok=True)
with open(os.path.expanduser("~/.config/adr/config.toml"), "w") as f:
f.write(f"""[adr.cache.stores]
file = {{ driver = "file", path = "{cache_path}" }}
""")
# Get the commits DB.
if db.is_old_version(
repository.COMMITS_DB) or not db.exists(repository.COMMITS_DB):
db.download(repository.COMMITS_DB, force=True)
# We'll use the past TRAINING_MONTHS months only for training the model,
# but we use 3 months more than that to calculate the failure statistics.
subprocess.run(
[
"run-adr",
"ahal/ci-recipes",
"recipe",
"-o",
os.path.abspath("push_data.json"),
"-f",
"json",
"push_data",
"--",
"--from",
f"today-{TRAINING_MONTHS + 3}month",
"--to",
"today-2day",
"--branch",
"autoland",
],
check=True,
stdout=subprocess.
DEVNULL, # Redirect to /dev/null, as the logs are too big otherwise.
)
HISTORY_DATE_START = datetime.now() - relativedelta(
months=TRAINING_MONTHS)
with open("push_data.json", "r") as f:
data = json.load(f)
push_data = {}
for row in data[1:]:
# Revision -> (all tasks, possible regressions, likely regressions)
push_data[row[0]] = (row[1], row[2], row[3])
HISTORICAL_TIMESPAN = 56
past_failures = {}
def get_past_failures(task, push_num):
if task not in past_failures:
past_failures[task] = repository.exp_queue(
push_num, HISTORICAL_TIMESPAN + 1, 0)
return past_failures[task][push_num]
def generate_data():
commits_with_data = set()
saved_nodes = set()
push_num = 0
for commit_data in tqdm(repository.get_commits()):
node = commit_data["node"]
if node not in push_data:
continue
commits_with_data.add(node)
commit_push_data = push_data[node]
for task in commit_push_data[0]:
if not any(task.startswith(j) for j in JOBS_TO_CONSIDER):
continue
total_failures = get_past_failures(task, push_num)
past_7_pushes_failures = total_failures - get_past_failures(
task, push_num - 7)
past_14_pushes_failures = total_failures - get_past_failures(
task, push_num - 14)
past_28_pushes_failures = total_failures - get_past_failures(
task, push_num - 28)
past_56_pushes_failures = total_failures - get_past_failures(
task, push_num - 56)
pushdate = dateutil.parser.parse(commit_data["pushdate"])
if pushdate > HISTORY_DATE_START:
saved_nodes.add(node)
yield {
"rev": node,
"name": task,
"failures": total_failures,
"failures_past_7_pushes": past_7_pushes_failures,
"failures_past_14_pushes": past_14_pushes_failures,
"failures_past_28_pushes": past_28_pushes_failures,
"failures_past_56_pushes": past_56_pushes_failures,
"is_possible_regression": task
in commit_push_data[1],
"is_likely_regression": task
in commit_push_data[2],
}
if task in commit_push_data[1] or task in commit_push_data[
2]:
past_failures[task][push_num] = total_failures + 1
push_num += 1
logger.info(f"push data nodes: {len(push_data)}")
logger.info(
f"commits linked to push data: {len(commits_with_data)}")
logger.info(f"saved push data nodes: {len(saved_nodes)}")
db.write(test_scheduling.TEST_SCHEDULING_DB, generate_data())
zstd_compress(test_scheduling.TEST_SCHEDULING_DB)
with tarfile.open("data/adr_cache.tar.xz", "w:xz") as tar:
tar.add("data/adr_cache")
def generate_test_scheduling_history(self, granularity):
push_data_path = f"push_data_{granularity}.json"
updated = download_check_etag(
test_scheduling.PUSH_DATA_URL.format(granularity=granularity))
if updated:
zstd_decompress(push_data_path)
os.remove(f"{push_data_path}.zst")
assert os.path.exists(
push_data_path), "Decompressed push data file exists"
# Get the commits DB.
assert db.download(repository.COMMITS_DB)
HISTORY_DATE_START = datetime.now() - relativedelta(
months=TRAINING_MONTHS[granularity])
if granularity == "label":
test_scheduling_db = test_scheduling.TEST_LABEL_SCHEDULING_DB
past_failures_db = os.path.join(
"data", test_scheduling.PAST_FAILURES_LABEL_DB)
failing_together_db = os.path.join(
"data", test_scheduling.FAILING_TOGETHER_LABEL_DB)
elif granularity == "group":
test_scheduling_db = test_scheduling.TEST_GROUP_SCHEDULING_DB
past_failures_db = os.path.join(
"data", test_scheduling.PAST_FAILURES_GROUP_DB)
touched_together_db = os.path.join(
"data", test_scheduling.TOUCHED_TOGETHER_DB)
db.download(test_scheduling_db, support_files_too=True)
last_node = None
for revs, _ in test_scheduling.get_test_scheduling_history(
granularity):
last_node = revs[0]
def generate_failing_together_probabilities(push_data):
# TODO: we should consider the probabilities of `task1 failure -> task2 failure` and
# `task2 failure -> task1 failure` separately, as they could be different.
count_runs = collections.Counter()
count_single_failures = collections.Counter()
count_both_failures = collections.Counter()
for revisions, tasks, likely_regressions, candidate_regressions in tqdm(
push_data):
failures = set(likely_regressions + candidate_regressions)
all_tasks = list(set(tasks) | failures)
for task1, task2 in itertools.combinations(
sorted(all_tasks), 2):
count_runs[(task1, task2)] += 1
if task1 in failures:
if task2 in failures:
count_both_failures[(task1, task2)] += 1
else:
count_single_failures[(task1, task2)] += 1
elif task2 in failures:
count_single_failures[(task1, task2)] += 1
stats = {}
skipped = 0
for couple, run_count in count_runs.most_common():
failure_count = count_both_failures[couple]
support = failure_count / run_count
if support < 1 / 700:
skipped += 1
continue
if failure_count != 0:
confidence = failure_count / (
count_single_failures[couple] + failure_count)
else:
confidence = 0.0
stats[couple] = (support, confidence)
logger.info(
f"{skipped} couples skipped because their support was too low")
logger.info(
"Redundancies with the highest support and confidence:")
for couple, (support,
confidence) in sorted(stats.items(),
key=lambda k:
(-k[1][1], -k[1][0]))[:7]:
failure_count = count_both_failures[couple]
run_count = count_runs[couple]
logger.info(
f"{couple[0]} - {couple[1]} redundancy confidence {confidence}, support {support} ({failure_count} over {run_count})."
)
logger.info(
"Redundancies with the highest confidence and lowest support:")
for couple, (support,
confidence) in sorted(stats.items(),
key=lambda k:
(-k[1][1], k[1][0]))[:7]:
failure_count = count_both_failures[couple]
run_count = count_runs[couple]
logger.info(
f"{couple[0]} - {couple[1]} redundancy confidence {confidence}, support {support} ({failure_count} over {run_count})."
)
failing_together = test_scheduling.get_failing_together_db()
count_redundancies = collections.Counter()
for couple, (support, confidence) in stats.items():
if confidence == 1.0:
count_redundancies["==100%"] += 1
if confidence > 0.9:
count_redundancies[">=90%"] += 1
if confidence > 0.8:
count_redundancies[">=80%"] += 1
if confidence > 0.7:
count_redundancies[">=70%"] += 1
if confidence < 0.7:
continue
failing_together[f"{couple[0]}${couple[1]}".encode(
"utf-8")] = struct.pack("ff", support, confidence)
for percentage, count in count_redundancies.most_common():
logger.info(f"{count} with {percentage} confidence")
test_scheduling.close_failing_together_db()
def generate_all_data():
past_failures = test_scheduling.get_past_failures(granularity)
push_num = past_failures[
"push_num"] if "push_num" in past_failures else 0
# We can start once we get to the last revision we added in the previous run.
can_start = True if last_node is None else False
commit_map = {}
for commit_data in tqdm(repository.get_commits()):
if not can_start:
if last_node == commit_data["node"]:
can_start = True
continue
commit_map[commit_data["node"]] = commit_data
with open(push_data_path, "r") as f:
push_data = json.load(f)
logger.info(f"push data nodes: {len(push_data)}")
if granularity == "label":
push_data = [(
revisions,
rename_tasks(push_tasks),
rename_tasks(possible_regressions),
rename_tasks(likely_regressions),
) for revisions, push_tasks, possible_regressions,
likely_regressions in push_data]
# In the last 14 pushes, we definitely run all possible runnables.
all_runnables_set = set(
sum((push_runnables
for _, push_runnables, _, _ in push_data[-14:]), []))
# Filter runnables we don't need.
all_runnables = filter_runnables(list(all_runnables_set),
all_runnables_set, granularity)
all_runnables_set = set(all_runnables_set)
logger.info(
f"{len(all_runnables_set)} runnables run in the last 14 pushes"
)
push_data = [(
revisions,
filter_runnables(push_tasks, all_runnables_set, granularity),
filter_runnables(possible_regressions, all_runnables_set,
granularity),
filter_runnables(likely_regressions, all_runnables_set,
granularity),
) for revisions, push_tasks, possible_regressions,
likely_regressions in push_data]
if granularity == "label":
generate_failing_together_probabilities(push_data)
# Store all runnables in the past_failures DB so it can be used in the evaluation phase.
past_failures["all_runnables"] = all_runnables
# XXX: Should we recreate the DB from scratch if the previous all_runnables are not the
# same as the current ones?
saved_nodes = set()
skipped_no_commits = 0
skipped_too_big_commits = 0
skipped_no_runnables = 0
# We can start once we get to the last revision we added in the previous run.
can_start = True if last_node is None else False
if granularity == "group":
update_touched_together_gen = test_scheduling.update_touched_together(
)
next(update_touched_together_gen)
for i in tqdm(range(len(push_data))):
(
revisions,
push_runnables,
possible_regressions,
likely_regressions,
) = push_data.pop(0)
if not can_start:
if last_node == revisions[0]:
can_start = True
continue
push_num += 1
# XXX: Some commits are skipped in the repository mining, e.g. merges and backouts. Maybe we should not skip them.
commits = tuple(
commit_map.pop(revision) for revision in revisions
if revision in commit_map)
if len(commits) == 0:
skipped_no_commits += 1
continue
merged_commits = commit_features.merge_commits(commits)
# XXX: For now, skip commits which are too large.
# In the future we can either:
# - Improve shelve perf and go back to consider all files;
# - Consider only files which appear with a given frequency, like the "files" feature in commit_features;
# - Keep a limit of number of files.
if len(merged_commits["files"]) > 50:
skipped_too_big_commits += 1
continue
# If we considered all_runnables, we'd generate a huge amount of data.
# We consider only the runnables which run in this push, and the possible and likely regressions
# from this push. We can't consider all runnables because we can't be sure that a task that didn't
# run on a push would have been successful.
runnables_to_consider = list(
set(push_runnables + possible_regressions +
likely_regressions))
if len(runnables_to_consider) == 0:
skipped_no_runnables += 1
continue
# Sync DB every 250 pushes, so we cleanup the shelve cache (we'd run OOM otherwise!).
if i % 250 == 0:
past_failures.sync()
pushdate = dateutil.parser.parse(merged_commits["pushdate"])
if granularity == "group":
update_touched_together_gen.send(commits[0]["node"])
result = {
"revs": revisions,
"data": [],
}
for data in test_scheduling.generate_data(
past_failures,
merged_commits,
push_num,
runnables_to_consider,
possible_regressions,
likely_regressions,
):
if pushdate > HISTORY_DATE_START:
result["data"].append(data)
if pushdate > HISTORY_DATE_START:
saved_nodes.add(i)
yield result
if granularity == "group":
try:
update_touched_together_gen.send(None)
except StopIteration:
pass
logger.info(f"saved push data nodes: {len(saved_nodes)}")
logger.info(f"skipped {skipped_no_commits} (no commits in our DB)")
logger.info(f"skipped {skipped_too_big_commits} (too big commits)")
logger.info(
f"skipped {skipped_no_runnables} (no interesting runnables)")
past_failures["push_num"] = push_num
past_failures.close()
db.append(test_scheduling_db, generate_all_data())
zstd_compress(test_scheduling_db)
create_tar_zst(past_failures_db)
if granularity == "group":
create_tar_zst(touched_together_db)
if granularity == "label":
create_tar_zst(failing_together_db)
def generate_push_data(
self, pushes: Tuple[mozci.push.Push, ...], granularity: str
) -> None:
from_date = get_from_date(granularity)
pushes = tuple(
push for push in pushes if datetime.utcfromtimestamp(push.date) >= from_date
)
if granularity == "label":
push_data_db = test_scheduling.PUSH_DATA_LABEL_DB
elif granularity == "group":
push_data_db = test_scheduling.PUSH_DATA_GROUP_DB
elif granularity == "config_group":
push_data_db = test_scheduling.PUSH_DATA_CONFIG_GROUP_DB
def cache_key(push: mozci.push.Push) -> str:
return f"push_data.{granularity}.{push.rev}"
def generate(
futures: List[concurrent.futures.Future],
) -> Generator[PushResult, None, None]:
num_cached = 0
num_pushes = len(pushes)
# Regenerating a large amount of data when we update the mozci regression detection
# algorithm is currently pretty slow, so we only regenerate 1000 pushes whenever we
# run.
to_regenerate = 1000
for push in tqdm(pushes):
cached = futures.pop(0).result()
semaphore.release()
if cached and to_regenerate > 0:
value, mozci_version = cached
# Regenerate results which were generated when we were not cleaning
# up WPT groups.
if granularity == "group" and any(
runnable.startswith("/") for runnable in value[1]
):
cached = None
to_regenerate -= 1
# Regenerate results which were generated when we didn't get a correct
# configuration for test-verify tasks.
elif granularity == "config_group" and any(
"test-verify" in runnable[0] for runnable in value[1]
):
cached = None
to_regenerate -= 1
"""# Regenerate results which were generated with an older version of mozci.
elif mozci_version != MOZCI_VERSION:
cached = None
to_regenerate -= 1"""
if cached:
num_cached += 1
value, mozci_version = cached
yield value
else:
logger.info(f"Analyzing {push.rev} at the {granularity} level...")
key = cache_key(push)
try:
if granularity == "label":
runnables = push.task_labels
elif granularity == "group":
runnables = push.group_summaries.keys()
elif granularity == "config_group":
runnables = push.config_group_summaries.keys()
value = (
push.revs,
tuple(runnables),
tuple(push.get_possible_regressions(granularity)),
tuple(push.get_likely_regressions(granularity)),
)
adr.config.cache.put(
key,
(value, MOZCI_VERSION),
adr.config["cache"]["retention"],
)
yield value
except adr.errors.MissingDataError:
logger.warning(
f"Tasks for push {push.rev} can't be found on ActiveData"
)
except Exception:
traceback.print_exc()
logger.info(f"{num_cached} pushes were already cached out of {num_pushes}")
semaphore = threading.BoundedSemaphore(256)
def retrieve_from_cache(push):
semaphore.acquire()
return adr.config.cache.get(cache_key(push))
with concurrent.futures.ThreadPoolExecutor() as executor:
futures = [executor.submit(retrieve_from_cache, push) for push in pushes]
try:
db.write(push_data_db, generate(futures))
except Exception:
for f in futures:
f.cancel()
try:
semaphore.release()
except ValueError:
continue
raise
zstd_compress(push_data_db)
def generate_push_data(self, runnable):
def upload_adr_cache():
cache_path = os.path.splitext(ADR_CACHE_DB)[0]
assert os.path.abspath(adr.config["cache"]["stores"]["file"]
["path"]) == os.path.abspath(cache_path)
with open_tar_zst(f"{ADR_CACHE_DB}.zst") as tar:
tar.add(cache_path)
db.upload(ADR_CACHE_DB)
# We keep in the cache the fact that we failed to analyze a push for 10
# days, so if we re-run often we don't retry the same pushes many times.
MISSING_CACHE_RETENTION = 10 * 24 * 60
# We'll use the past TRAINING_MONTHS months only for training the model,
# but we use half TRAINING_MONTHS months more than that to calculate the
# failure statistics.
from_months = TRAINING_MONTHS[runnable] + math.floor(
TRAINING_MONTHS[runnable] / 2)
pushes = mozci.push.make_push_objects(
from_date=f"today-{from_months}month",
to_date="today-3day",
branch="autoland",
)
start_time = time.monotonic()
num_cached = 0
push_data = []
for push in tqdm(pushes):
key = f"push_data.{runnable}.{push.rev}"
logger.info(f"Analyzing {push.rev} at the {runnable} level...")
if adr.config.cache.has(key):
num_cached += 1
cached = adr.config.cache.get(key)
if cached:
# XXX: We have to support items in the cache that were added
# before the mozci version was stored. We can drop the if
# when all items have been switched over.
value = cached[0] if isinstance(cached, tuple) else cached
push_data.append(value)
else:
try:
if runnable == "label":
runnables = push.task_labels
elif runnable == "group":
runnables = push.group_summaries.keys()
value = [
push.revs,
list(runnables),
list(push.get_possible_regressions(runnable)),
list(push.get_likely_regressions(runnable)),
]
push_data.append(value)
adr.config.cache.forever(key, (value, MOZCI_VERSION))
except adr.errors.MissingDataError:
logger.warning(
f"Tasks for push {push.rev} can't be found on ActiveData"
)
adr.config.cache.put(key, (), MISSING_CACHE_RETENTION)
except Exception:
traceback.print_exc()
adr.config.cache.put(key, (), MISSING_CACHE_RETENTION)
if time.monotonic() - start_time >= 3600:
upload_adr_cache()
start_time = time.monotonic()
logger.info(
f"{num_cached} pushes were already cached out of {len(pushes)}")
upload_adr_cache()
with open(f"push_data_{runnable}.json", "w") as f:
json.dump(push_data, f)
zstd_compress(f"push_data_{runnable}.json")
def generate_test_scheduling_history(self):
if not os.path.exists("push_data.json"):
download_check_etag(PUSH_DATA_URL, "push_data.json.zst")
zstd_decompress("push_data.json")
assert os.path.exists(
"push_data.json"
), "Decompressed push data file exists"
# Get the commits DB.
if db.is_old_version(repository.COMMITS_DB) or not db.exists(
repository.COMMITS_DB
):
db.download(repository.COMMITS_DB, force=True)
HISTORY_DATE_START = datetime.now() - relativedelta(months=TRAINING_MONTHS)
with open("push_data.json", "r") as f:
data = json.load(f)
push_data = {}
for row in data[1:]:
# Revision -> (all tasks, possible regressions, likely regressions)
push_data[row[0]] = (row[1], row[2], row[3])
HISTORICAL_TIMESPAN = 56
if not db.is_old_version(test_scheduling.TEST_SCHEDULING_DB):
db.download(test_scheduling.TEST_SCHEDULING_DB, support_files_too=True)
for test_data in test_scheduling.get_test_scheduling_history():
pass
last_node = test_data["rev"]
else:
last_node = None
try:
with open("data/past_failures.pickle", "rb") as f:
past_failures, push_num = pickle.load(f)
except FileNotFoundError:
past_failures = {}
push_num = 0
def get_and_update_past_failures(type_, task, items, push_num, is_regression):
if type_ not in past_failures:
past_failures[type_] = {}
if task not in past_failures[type_]:
past_failures[type_][task] = {}
values_total = []
values_prev_7 = []
values_prev_14 = []
values_prev_28 = []
values_prev_56 = []
for item in items:
if item not in past_failures[type_][task]:
past_failures[type_][task][item] = ExpQueue(
push_num, HISTORICAL_TIMESPAN + 1, 0
)
value = past_failures[type_][task][item][push_num]
values_total.append(value)
values_prev_7.append(
value - past_failures[type_][task][item][push_num - 7]
)
values_prev_14.append(
value - past_failures[type_][task][item][push_num - 14]
)
values_prev_28.append(
value - past_failures[type_][task][item][push_num - 28]
)
values_prev_56.append(
value - past_failures[type_][task][item][push_num - 56]
)
if is_regression:
past_failures[type_][task][item][push_num] = value + 1
return (
sum(values_total),
sum(values_prev_7),
sum(values_prev_14),
sum(values_prev_28),
sum(values_prev_56),
)
def generate_data():
nonlocal push_num
commits_with_data = set()
saved_nodes = set()
# We can start once we get to the last revision we added in the previous run.
can_start = True if last_node is None else False
for commit_data in tqdm(repository.get_commits()):
node = commit_data["node"]
if node == last_node:
can_start = True
continue
if not can_start:
continue
if node not in push_data:
continue
commits_with_data.add(node)
commit_push_data = push_data[node]
for task in commit_push_data[0]:
if not any(task.startswith(j) for j in JOBS_TO_CONSIDER):
continue
is_regression = (
task in commit_push_data[1] or task in commit_push_data[2]
)
total_failures, past_7_pushes_failures, past_14_pushes_failures, past_28_pushes_failures, past_56_pushes_failures = get_and_update_past_failures(
"all", task, ["all"], push_num, is_regression
)
total_types_failures, past_7_pushes_types_failures, past_14_pushes_types_failures, past_28_pushes_types_failures, past_56_pushes_types_failures = get_and_update_past_failures(
"type", task, commit_data["types"], push_num, is_regression
)
total_files_failures, past_7_pushes_files_failures, past_14_pushes_files_failures, past_28_pushes_files_failures, past_56_pushes_files_failures = get_and_update_past_failures(
"file", task, commit_data["files"], push_num, is_regression
)
total_directories_failures, past_7_pushes_directories_failures, past_14_pushes_directories_failures, past_28_pushes_directories_failures, past_56_pushes_directories_failures = get_and_update_past_failures(
"directory",
task,
commit_data["directories"],
push_num,
is_regression,
)
total_components_failures, past_7_pushes_components_failures, past_14_pushes_components_failures, past_28_pushes_components_failures, past_56_pushes_components_failures = get_and_update_past_failures(
"component",
task,
commit_data["components"],
push_num,
is_regression,
)
pushdate = dateutil.parser.parse(commit_data["pushdate"])
if pushdate > HISTORY_DATE_START:
saved_nodes.add(node)
yield {
"rev": node,
"name": task,
"failures": total_failures,
"failures_past_7_pushes": past_7_pushes_failures,
"failures_past_14_pushes": past_14_pushes_failures,
"failures_past_28_pushes": past_28_pushes_failures,
"failures_past_56_pushes": past_56_pushes_failures,
"failures_in_types": total_types_failures,
"failures_past_7_pushes_in_types": past_7_pushes_types_failures,
"failures_past_14_pushes_in_types": past_14_pushes_types_failures,
"failures_past_28_pushes_in_types": past_28_pushes_types_failures,
"failures_past_56_pushes_in_types": past_56_pushes_types_failures,
"failures_in_files": total_files_failures,
"failures_past_7_pushes_in_files": past_7_pushes_files_failures,
"failures_past_14_pushes_in_files": past_14_pushes_files_failures,
"failures_past_28_pushes_in_files": past_28_pushes_files_failures,
"failures_past_56_pushes_in_files": past_56_pushes_files_failures,
"failures_in_directories": total_directories_failures,
"failures_past_7_pushes_in_directories": past_7_pushes_directories_failures,
"failures_past_14_pushes_in_directories": past_14_pushes_directories_failures,
"failures_past_28_pushes_in_directories": past_28_pushes_directories_failures,
"failures_past_56_pushes_in_directories": past_56_pushes_directories_failures,
"failures_in_components": total_components_failures,
"failures_past_7_pushes_in_components": past_7_pushes_components_failures,
"failures_past_14_pushes_in_components": past_14_pushes_components_failures,
"failures_past_28_pushes_in_components": past_28_pushes_components_failures,
"failures_past_56_pushes_in_components": past_56_pushes_components_failures,
"is_possible_regression": task in commit_push_data[1],
"is_likely_regression": task in commit_push_data[2],
}
push_num += 1
logger.info(f"push data nodes: {len(push_data)}")
logger.info(f"commits linked to push data: {len(commits_with_data)}")
logger.info(f"saved push data nodes: {len(saved_nodes)}")
db.append(test_scheduling.TEST_SCHEDULING_DB, generate_data())
zstd_compress(test_scheduling.TEST_SCHEDULING_DB)
with open("data/past_failures.pickle", "wb") as f:
pickle.dump((past_failures, push_num), f, protocol=pickle.HIGHEST_PROTOCOL)
zstd_compress("data/past_failures.pickle")
