def prerun_check(project_id, run_id, parent_span=None):
log = logger.bind(pid=project_id, run_id=run_id)
log.debug("Sanity check that we need to compute run")
with DBConn() as conn:
if not check_project_exists(conn, project_id):
log.debug("Project not found. Skipping")
raise ProjectDeleted(project_id)
res = get_run(conn, run_id)
if res is None:
log.debug(f"Run not found. Skipping")
raise RunDeleted(run_id)
try:
state = get_run_state_for_update(conn, run_id)
except psycopg2.OperationalError:
log.warning("Run started in another task. Skipping this race.")
return
if state in {'running', 'completed', 'error'}:
log.warning("Run already started. Skipping")
return
log.debug("Setting run as in progress")
update_run_set_started(conn, run_id)
log.debug("Getting dp ids for compute similarity task")
dp_ids = get_dataprovider_ids(conn, project_id)
log.debug("Data providers: {}".format(dp_ids))
create_comparison_jobs.delay(project_id, run_id,
prerun_check.get_serialized_span())
log.info("CLK similarity computation scheduled")
def save_and_permute(similarity_result, project_id, run_id, parent_span):
log = logger.bind(pid=project_id, run_id=run_id)
log.debug("Saving and possibly permuting data")
mapping = similarity_result['mapping']
# Note Postgres requires JSON object keys to be strings
# Celery actually converts the json arguments in the same way
with DBConn() as db:
result_type = get_project_column(db, project_id, 'result_type')
# Just save the raw "mapping"
log.debug("Saving the resulting map data to the db")
result_id = insert_mapping_result(db, run_id, mapping)
dp_ids = get_dataprovider_ids(db, project_id)
log.info("Mapping result saved to db with result id {}".format(result_id))
if result_type == "permutations":
log.debug("Submitting job to permute mapping")
permute_mapping_data.apply_async(
(project_id, run_id,
similarity_result['lenf1'], similarity_result['lenf2'],
save_and_permute.get_serialized_span()))
else:
log.debug("Mark mapping job as complete")
mark_run_complete.delay(run_id, save_and_permute.get_serialized_span())
# Post similarity computation cleanup
log.debug("Removing clk filters from redis cache")
for dp_id in dp_ids:
cache.remove_from_cache(dp_id)
calculate_comparison_rate.delay()
def create_comparison_jobs(project_id, run_id, parent_span=None):
"""Schedule all the entity comparisons as sub tasks for a run.
At a high level this task:
- checks if the project and run have been deleted and if so aborts.
- retrieves metadata: the number and size of the datasets, the encoding size,
and the number and size of blocks.
- splits the work into independent "chunks" and schedules them to run in celery
- schedules the follow up task to run after all the comparisons have been computed.
"""
log = logger.bind(pid=project_id, run_id=run_id)
current_span = create_comparison_jobs.span
with DBConn() as conn:
check_run_active(conn, project_id, run_id)
dp_ids = get_dataprovider_ids(conn, project_id)
number_of_datasets = len(dp_ids)
assert number_of_datasets >= 2, "Expected at least 2 data providers"
log.info(f"Scheduling comparison of CLKs from data provider ids: "
f"{', '.join(map(str, dp_ids))}")
# Retrieve required metadata
dataset_sizes, dp_block_sizes = _retrieve_blocked_dataset_sizes(
conn, project_id, dp_ids)
log.info("Finding blocks in common between dataproviders")
common_blocks = _get_common_blocks(dp_block_sizes, dp_ids)
# We pass the encoding_size and threshold to the comparison tasks to minimize their db lookups
encoding_size = get_project_encoding_size(conn, project_id)
threshold = get_run(conn, run_id)['threshold']
log.debug("Chunking computation task")
# Create "chunks" of comparisons
chunks = _create_work_chunks(common_blocks, dp_block_sizes, dp_ids, log)
log.info(f"Chunking into {len(chunks)} computation tasks")
current_span.log_kv({
"event": "chunking",
'num_chunks': len(chunks),
'dataset-sizes': dataset_sizes
})
span_serialized = create_comparison_jobs.get_serialized_span()
# Prepare the Celery Chord that will compute all the similarity scores:
scoring_tasks = [
compute_filter_similarity.si(chunk_info, project_id, run_id, threshold,
encoding_size, span_serialized)
for chunk_info in chunks
]
if len(scoring_tasks) == 1:
scoring_tasks.append(celery_bug_fix.si())
callback_task = aggregate_comparisons.s(
project_id=project_id, run_id=run_id,
parent_span=span_serialized).on_error(
run_failed_handler.s(run_id=run_id))
log.info(f"Scheduling comparison tasks")
future = chord(scoring_tasks)(callback_task)
def aggregate_comparisons(similarity_result_files, project_id, run_id, parent_span=None):
log = logger.bind(pid=project_id, run_id=run_id)
if similarity_result_files is None: return
mc = connect_to_object_store()
files = []
data_size = 0
for num, filename in similarity_result_files:
if num > 0:
files.append(filename)
data_size += mc.stat_object(Config.MINIO_BUCKET, filename).size
log.debug("Aggregating result chunks from {} files, total size: {}".format(
len(files), fmt_bytes(data_size)))
result_file_stream_generator = (mc.get_object(Config.MINIO_BUCKET, result_filename) for result_filename in files)
log.info("Similarity score results are {}".format(fmt_bytes(data_size)))
result_stream = chain_streams(result_file_stream_generator)
with DBConn() as db:
result_type = get_project_column(db, project_id, 'result_type')
# Note: Storing the similarity scores for all result types
result_filename = store_similarity_scores(result_stream, run_id, data_size, db)
if result_type == "similarity_scores":
# Post similarity computation cleanup
dp_ids = get_dataprovider_ids(db, project_id)
else:
# we promote the run to the next stage
progress_stage(db, run_id)
lenf1, lenf2 = get_project_dataset_sizes(db, project_id)
# DB now committed, we can fire off tasks that depend on the new db state
if result_type == "similarity_scores":
log.info("Deleting intermediate similarity score files from object store")
mc.remove_objects(Config.MINIO_BUCKET, files)
log.debug("Removing clk filters from redis cache")
remove_from_cache(dp_ids[0])
remove_from_cache(dp_ids[1])
# Complete the run
log.info("Marking run as complete")
mark_run_complete.delay(run_id, aggregate_comparisons.get_serialized_span())
else:
solver_task.delay(result_filename, project_id, run_id, lenf1, lenf2, aggregate_comparisons.get_serialized_span())
def save_and_permute(similarity_result, project_id, run_id, parent_span):
log = logger.bind(pid=project_id, run_id=run_id)
log.debug("Saving and possibly permuting data")
groups = similarity_result['groups']
# Note Postgres requires JSON object keys to be strings
# Celery actually converts the json arguments in the same way
with DBConn() as db:
result_type = get_project_column(db, project_id, 'result_type')
if result_type == "groups":
# Save the raw groups
log.debug("Saving the groups in the DB")
result_id = insert_mapping_result(db, run_id, groups)
else:
# Turn groups into mapping and save that
log.debug("Turning groups into mapping")
mapping = groups_to_mapping(groups)
log.debug("Saving mappuing in the DB")
result_id = insert_mapping_result(db, run_id, mapping)
dp_ids = get_dataprovider_ids(db, project_id)
log.info("Result saved to db with result id {}".format(result_id))
if result_type == "permutations":
log.debug("Submitting job to permute mapping")
dataset0_size, dataset1_size = similarity_result['datasetSizes']
permute_mapping_data.apply_async(
(project_id, run_id, dataset0_size, dataset1_size,
save_and_permute.get_serialized_span()))
else:
log.debug("Mark job as complete")
mark_run_complete.delay(run_id, save_and_permute.get_serialized_span())
# Post similarity computation cleanup
log.debug("Removing clk filters from redis cache")
for dp_id in dp_ids:
cache.remove_from_cache(dp_id)
calculate_comparison_rate.delay()
def aggregate_comparisons(similarity_result_files,
project_id,
run_id,
parent_span=None):
log = logger.bind(pid=project_id, run_id=run_id)
if similarity_result_files is None:
raise TypeError("Inappropriate argument type - missing results files.")
files = []
for res in similarity_result_files:
if res is None:
log.warning(
"Missing results during aggregation. Stopping processing.")
raise TypeError(
"Inappropriate argument type - results missing at aggregation step."
)
num, filesize, filename = res
if num:
assert filesize is not None
assert filename is not None
files.append((num, filesize, filename))
else:
assert filesize is None
assert filename is None
heapq.heapify(files)
log.debug(f"Aggregating result chunks from {len(files)} files, "
f"total size: {sum(map(operator.itemgetter(1), files))}")
mc = connect_to_object_store()
while len(files) > 1:
file0 = heapq.heappop(files)
file1 = heapq.heappop(files)
merged_file = _merge_files(mc, log, file0, file1)
heapq.heappush(files, merged_file)
if not files:
# No results. Let's chuck in an empty file.
empty_file = _put_placeholder_empty_file(mc, log)
files.append(empty_file)
(merged_num, merged_filesize, merged_filename), = files
log.info(f"Similarity score results in {merged_filename} in bucket "
f"{Config.MINIO_BUCKET} take up {merged_filesize} bytes.")
with DBConn() as db:
result_type = get_project_column(db, project_id, 'result_type')
result_id = insert_similarity_score_file(db, run_id, merged_filename)
log.debug(f"Saved path to similarity scores file to db with id "
f"{result_id}")
if result_type == "similarity_scores":
# Post similarity computation cleanup
dp_ids = get_dataprovider_ids(db, project_id)
else:
# we promote the run to the next stage
progress_stage(db, run_id)
dataset_sizes = get_project_dataset_sizes(db, project_id)
# DB now committed, we can fire off tasks that depend on the new db state
if result_type == "similarity_scores":
log.debug("Removing clk filters from redis cache")
for dp_id in dp_ids:
remove_from_cache(dp_id)
# Complete the run
log.info("Marking run as complete")
mark_run_complete.delay(run_id,
aggregate_comparisons.get_serialized_span())
else:
solver_task.delay(merged_filename, project_id, run_id, dataset_sizes,
aggregate_comparisons.get_serialized_span())
def permute_mapping_data(project_id, run_id, len_filters1, len_filters2,
parent_span):
"""
Task which will create a permutation after a mapping has been completed.
:param project_id: The project resource id
:param run_id: The run id
:param len_filters1:
:param len_filters2:
"""
log = logger.bind(pid=project_id, run_id=run_id)
with DBConn() as conn:
mapping_str = get_run_result(conn, run_id)
# Convert to int: int
mapping = {int(k): int(mapping_str[k]) for k in mapping_str}
log.info("Creating random permutations")
log.debug(
"Entities in dataset A: {}, Entities in dataset B: {}".format(
len_filters1, len_filters2))
"""
Pack all the entities that match in the **same** random locations in both permutations.
Then fill in all the gaps!
Dictionaries first, then converted to lists.
"""
smaller_dataset_size = min(len_filters1, len_filters2)
log.debug("Smaller dataset size is {}".format(smaller_dataset_size))
number_in_common = len(mapping)
a_permutation = {} # Should be length of filters1
b_permutation = {} # length of filters2
# By default mark all rows as NOT included in the mask
mask = {i: False for i in range(smaller_dataset_size)}
# start with all the possible indexes
remaining_new_indexes = list(range(smaller_dataset_size))
log.info("Shuffling indices for matched entities")
random.shuffle(remaining_new_indexes)
log.info("Assigning random indexes for {} matched entities".format(
number_in_common))
for mapping_number, a_index in enumerate(mapping):
b_index = mapping[a_index]
# Choose the index in the new mapping (randomly)
mapping_index = remaining_new_indexes[mapping_number]
a_permutation[a_index] = mapping_index
b_permutation[b_index] = mapping_index
# Mark the row included in the mask
mask[mapping_index] = True
remaining_new_indexes = set(remaining_new_indexes[number_in_common:])
log.info("Randomly adding all non matched entities")
# Note the a and b datasets could be of different size.
# At this point, both still have to use the remaining_new_indexes, and any
# indexes that go over the number_in_common
remaining_a_values = list(
set(range(smaller_dataset_size,
len_filters1)).union(remaining_new_indexes))
remaining_b_values = list(
set(range(smaller_dataset_size,
len_filters2)).union(remaining_new_indexes))
log.debug("Shuffle the remaining indices")
random.shuffle(remaining_a_values)
random.shuffle(remaining_b_values)
# For every element in a's permutation
for a_index in range(len_filters1):
# Check if it is not already present
if a_index not in a_permutation:
# This index isn't yet mapped
# choose and remove a random index from the extended list of those that remain
# note this "could" be the same row (a NOP 1-1 permutation)
mapping_index = remaining_a_values.pop()
a_permutation[a_index] = mapping_index
# For every eventual element in a's permutation
for b_index in range(len_filters2):
# Check if it is not already present
if b_index not in b_permutation:
# This index isn't yet mapped
# choose and remove a random index from the extended list of those that remain
# note this "could" be the same row (a NOP 1-1 permutation)
mapping_index = remaining_b_values.pop()
b_permutation[b_index] = mapping_index
log.debug("Completed creating new permutations for each party")
dp_ids = get_dataprovider_ids(conn, project_id)
for i, permutation in enumerate([a_permutation, b_permutation]):
# We convert here because celery and dicts with int keys don't play nice
perm_list = convert_mapping_to_list(permutation)
log.debug("Saving a permutation")
insert_permutation(conn, dp_ids[i], run_id, perm_list)
log.debug("Raw permutation data saved. Now saving raw mask")
# Convert the mask dict to a list of 0/1 ints
mask_list = convert_mapping_to_list(
{int(key): 1 if value else 0
for key, value in mask.items()})
log.debug("Saving the mask")
insert_permutation_mask(conn, project_id, run_id, mask_list)
log.info("Mask saved")
log.info("Committing database transaction")
mark_run_complete.delay(run_id, permute_mapping_data.get_serialized_span())
def create_comparison_jobs(project_id, run_id, parent_span=None):
log = logger.bind(pid=project_id, run_id=run_id)
with DBConn() as conn:
dp_ids = get_dataprovider_ids(conn, project_id)
assert len(dp_ids) >= 2, "Expected at least 2 data providers"
log.info(f"Starting comparison of CLKs from data provider ids: "
f"{', '.join(map(str, dp_ids))}")
current_span = create_comparison_jobs.span
if not check_project_exists(conn, project_id) or not check_run_exists(
conn, project_id, run_id):
log.info("Skipping as project or run not found in database.")
return
run_info = get_run(conn, run_id)
threshold = run_info['threshold']
dataset_sizes = get_project_dataset_sizes(conn, project_id)
if len(dataset_sizes) < 2:
log.warning("Unexpected number of dataset sizes in db. Stopping")
update_run_mark_failure(conn, run_id)
return
encoding_size = get_project_encoding_size(conn, project_id)
log.info(f"Computing similarity for "
f"{' x '.join(map(str, dataset_sizes))} entities")
current_span.log_kv({"event": 'get-dataset-sizes'})
filters_object_filenames = tuple(
get_filter_metadata(conn, dp_id) for dp_id in dp_ids)
current_span.log_kv({"event": 'get-metadata'})
log.debug("Chunking computation task")
chunk_infos = tuple(
anonlink.concurrency.split_to_chunks(Config.CHUNK_SIZE_AIM,
dataset_sizes=dataset_sizes))
# Save filenames with chunk information.
for chunk_info in chunk_infos:
for chunk_dp_info in chunk_info:
chunk_dp_index = chunk_dp_info['datasetIndex']
chunk_dp_store_filename = filters_object_filenames[chunk_dp_index]
chunk_dp_info['storeFilename'] = chunk_dp_store_filename
log.info(f"Chunking into {len(chunk_infos)} computation tasks")
current_span.log_kv({"event": "chunking", 'num_chunks': len(chunk_infos)})
span_serialized = create_comparison_jobs.get_serialized_span()
# Prepare the Celery Chord that will compute all the similarity scores:
scoring_tasks = [
compute_filter_similarity.si(chunk_info, project_id, run_id, threshold,
encoding_size, span_serialized)
for chunk_info in chunk_infos
]
if len(scoring_tasks) == 1:
scoring_tasks.append(celery_bug_fix.si())
callback_task = aggregate_comparisons.s(
project_id, run_id,
parent_span=span_serialized).on_error(on_chord_error.s(run_id=run_id))
future = chord(scoring_tasks)(callback_task)
def create_comparison_jobs(project_id, run_id, parent_span=None):
log = logger.bind(pid=project_id, run_id=run_id)
with DBConn() as conn:
dp_ids = get_dataprovider_ids(conn, project_id)
assert len(dp_ids) >= 2, "Expected at least 2 data providers"
log.info("Starting comparison of CLKs from data provider ids: {}, {}".format(dp_ids[0], dp_ids[1]))
current_span = create_comparison_jobs.span
if not check_project_exists(conn, project_id) or not check_run_exists(conn, project_id, run_id):
log.info("Skipping as project or run not found in database.")
return
run_info = get_run(conn, run_id)
threshold = run_info['threshold']
dataset_sizes = get_project_dataset_sizes(conn, project_id)
if len(dataset_sizes) < 2:
log.warning("Unexpected number of dataset sizes in db. Stopping")
update_run_mark_failure(conn, run_id)
return
else:
lenf1, lenf2 = dataset_sizes
encoding_size = get_project_encoding_size(conn, project_id)
size = lenf1 * lenf2
log.info("Computing similarity for {} x {} entities".format(lenf1, lenf2))
current_span.log_kv({"event": 'get-dataset-sizes'})
filters1_object_filename = get_filter_metadata(conn, dp_ids[0])
filters2_object_filename = get_filter_metadata(conn, dp_ids[1])
current_span.log_kv({"event": 'get-metadata'})
log.debug("Chunking computation task")
chunk_size = Config.get_task_chunk_size(size, threshold)
if chunk_size is None:
chunk_size = max(lenf1, lenf2)
log.info("Chunks will contain {} entities per task".format(chunk_size))
update_run_chunk(conn, project_id, chunk_size)
job_chunks = []
dp1_chunks = []
dp2_chunks = []
for chunk_start_index_dp1 in range(0, lenf1, chunk_size):
dp1_chunks.append(
(filters1_object_filename, chunk_start_index_dp1, min(chunk_start_index_dp1 + chunk_size, lenf1))
)
for chunk_start_index_dp2 in range(0, lenf2, chunk_size):
dp2_chunks.append(
(filters2_object_filename, chunk_start_index_dp2, min(chunk_start_index_dp2 + chunk_size, lenf2))
)
# Every chunk in dp1 has to be run against every chunk in dp2
for dp1_chunk in dp1_chunks:
for dp2_chunk in dp2_chunks:
job_chunks.append((dp1_chunk, dp2_chunk, ))
log.info("Chunking into {} computation tasks each with (at most) {} entities.".format(
len(job_chunks), chunk_size))
current_span.log_kv({"event": "chunking", "chunksize": chunk_size, 'num_chunks': len(job_chunks)})
span_serialized = create_comparison_jobs.get_serialized_span()
# Prepare the Celery Chord that will compute all the similarity scores:
scoring_tasks = [compute_filter_similarity.si(
chunk_dp1,
chunk_dp2,
project_id,
run_id,
threshold,
encoding_size,
span_serialized
) for chunk_dp1, chunk_dp2 in job_chunks]
if len(scoring_tasks) == 1:
scoring_tasks.append(celery_bug_fix.si())
callback_task = aggregate_comparisons.s(project_id, run_id, parent_span=span_serialized).on_error(
on_chord_error.s(run_id=run_id))
future = chord(scoring_tasks)(callback_task)