def get_levels_overview(group):
query = (Query("events", Entity("events")).set_select([
Column("primary_hash"),
Function("max", [Function("length", [Column("hierarchical_hashes")])],
"num_levels"),
_current_level_expr(group),
]).set_where(_get_group_filters(group)).set_groupby(
[Column("primary_hash")]))
res = snuba.raw_snql_query(
query, referrer="api.group_hashes_levels.get_levels_overview")
if not res["data"]:
raise NoEvents()
if len(res["data"]) > 1:
raise MergedIssues()
assert len(res["data"]) == 1
fields = res["data"][0]
if fields["num_levels"] <= 0:
raise NotHierarchical()
# TODO: Cache this if it takes too long. This is called from multiple
# places, grouping overview and then again in the new-issues endpoint.
return LevelsOverview(
current_level=fields["current_level"] - 1,
only_primary_hash=fields["primary_hash"],
num_levels=fields["num_levels"],
)
def _current_level_expr(group):
materialized_hashes = {
gh.hash
for gh in GroupHash.objects.filter(project=group.project, group=group)
}
# Evaluates to the index of the last hash that is in materialized_hashes,
# or 1 otherwise.
find_hash_expr = _construct_arraymax([1] + [ # type: ignore
Function("indexOf", [Column("hierarchical_hashes"), hash])
for hash in materialized_hashes
])
return Function("max", [find_hash_expr], "current_level")
def _get_hash_for_parent_level(group: Group, id: int,
levels_overview: LevelsOverview) -> str:
# If this is violated, there cannot be a 1:1 mapping between level and hash.
assert 0 <= id < levels_overview.current_level
# This cache never needs explicit invalidation because during every level
# change, the group ID changes.
#
# No idea if the query is slow, caching just because I can.
cache_key = f"group-parent-level-hash:{group.id}:{id}"
return_hash: str = cache.get(cache_key)
if return_hash is None:
query = (Query("events", Entity("events")).set_select([
Function("arrayElement", [Column("hierarchical_hashes"), id + 1],
"hash")
]).set_where(_get_group_filters(group)).set_limit(1))
return_hash: str = get_path(snuba.raw_snql_query(query), "data", 0,
"hash") # type: ignore
cache.set(cache_key, return_hash)
assert return_hash
return return_hash
def _get_full_hierarchical_hashes(group: Group, hash: str) -> Optional[Sequence[str]]:
query = (
Query("events", Entity("events"))
.set_select(
[
Column("hierarchical_hashes"),
]
)
.set_where(
_get_group_filters(group)
+ [
Condition(
Function(
"has",
[Column("hierarchical_hashes"), hash],
),
Op.EQ,
1,
),
]
)
)
data = snuba.raw_snql_query(query, referrer="group_split.get_full_hierarchical_hashes")["data"]
if not data:
return None
return data[0]["hierarchical_hashes"]
def _construct_arraymax(elements):
# XXX(markus): This is quite horrible but Snuba SDK does not allow us to do
# arrayMax([<other function call>, ...]), i.e. it does not allow function
# calls in array literals. So instead of arrayMax([1, 2, 3]) we do
# greatest(1, greatest(2, 3)).
assert elements
if len(elements) == 1:
return elements[0]
# Attempt to build well-balanced 'tree' of greatest() such that
# we don't run into ClickHouse recursion limits.
return Function(
"greatest",
[
_construct_arraymax(elements[:len(elements) // 2]),
_construct_arraymax(elements[len(elements) // 2:]),
],
)
def query(
self,
projects: Sequence[Project],
retention_window_start: Optional[datetime],
group_queryset: QuerySet,
environments: Sequence[Environment],
sort_by: str,
limit: int,
cursor: Optional[Cursor],
count_hits: bool,
paginator_options: Mapping[str, Any],
search_filters: Sequence[SearchFilter],
date_from: Optional[datetime],
date_to: Optional[datetime],
max_hits=None,
) -> CursorResult:
if not validate_cdc_search_filters(search_filters):
raise InvalidQueryForExecutor(
"Search filters invalid for this query executor")
start, end, retention_date = self.calculate_start_end(
retention_window_start, search_filters, date_from, date_to)
if start == retention_date and end == retention_date:
# Both `start` and `end` must have been trimmed to `retention_date`,
# so this entire search was against a time range that is outside of
# retention. We'll return empty results to maintain backwards compatibility
# with Django search (for now).
return self.empty_result
if start >= end:
# TODO: This maintains backwards compatibility with Django search, but
# in the future we should find a way to notify the user that their search
# is invalid.
return self.empty_result
e_event = self.entities["event"]
e_group = self.entities["group"]
where_conditions = [
Condition(Column("project_id", e_event), Op.IN,
[p.id for p in projects]),
Condition(Column("timestamp", e_event), Op.GTE, start),
Condition(Column("timestamp", e_event), Op.LT, end),
]
# TODO: This is still basically only handling status, handle this better once we introduce
# more conditions.
for search_filter in search_filters:
where_conditions.append(
Condition(Column(search_filter.key.name, e_group), Op.IN,
search_filter.value.raw_value))
if environments:
# TODO: Should this be handled via filter_keys, once we have a snql compatible version?
where_conditions.append(
Condition(Column("environment", e_event), Op.IN,
[e.name for e in environments]))
sort_func = self.aggregation_defs[self.sort_strategies[sort_by]]
having = []
if cursor is not None:
op = Op.GTE if cursor.is_prev else Op.LTE
having.append(Condition(sort_func, op, cursor.value))
query = Query(
"events",
match=Join([Relationship(e_event, "grouped", e_group)]),
select=[
Column("id", e_group),
replace(sort_func, alias="score"),
],
where=where_conditions,
groupby=[Column("id", e_group)],
having=having,
orderby=[OrderBy(sort_func, direction=Direction.DESC)],
limit=Limit(limit + 1),
)
data = snuba.raw_snql_query(
query, referrer="search.snuba.cdc_search.query")["data"]
hits_query = Query(
"events",
match=Join([Relationship(e_event, "grouped", e_group)]),
select=[
Function("uniq", [Column("id", e_group)], alias="count"),
],
where=where_conditions,
)
hits = None
if count_hits:
hits = snuba.raw_snql_query(
hits_query,
referrer="search.snuba.cdc_search.hits")["data"][0]["count"]
paginator_results = SequencePaginator(
[(row["score"], row["g.id"]) for row in data],
reverse=True,
**paginator_options,
).get_result(limit, cursor, known_hits=hits, max_hits=max_hits)
# We filter against `group_queryset` here so that we recheck all conditions in Postgres.
# Since replay between Postgres and Clickhouse can happen, we might get back results that
# have changed state in Postgres. By rechecking them we guarantee than any returned results
# have the correct state.
# TODO: This can result in us returning less than a full page of results, but shouldn't
# affect cursors. If we want to, we can iterate and query snuba until we manage to get a
# full page. In practice, this will likely only skip a couple of results at worst, and
# probably not be noticeable to the user, so holding off for now to reduce complexity.
groups = group_queryset.in_bulk(paginator_results.results)
paginator_results.results = [
groups[k] for k in paginator_results.results if k in groups
]
return paginator_results
class CdcPostgresSnubaQueryExecutor(PostgresSnubaQueryExecutor):
sort_strategies = {
"date": "last_seen",
"freq": "times_seen",
"new": "first_seen",
"priority": "priority",
"user": "******",
}
entities = {
"event": Entity("events", alias="e"),
"group": Entity("groupedmessage", alias="g"),
}
times_seen_aggregation = Function(
"ifNull",
[Function("count", [Column("group_id", entities["event"])]), 0])
first_seen_aggregation = Function(
"ifNull",
[
Function(
"multiply",
[
Function("toUInt64", [
Function("min",
[Column("timestamp", entities["event"])])
]),
1000,
],
),
0,
],
)
last_seen_aggregation = Function(
"ifNull",
[
Function(
"multiply",
[
Function("toUInt64", [
Function("max",
[Column("timestamp", entities["event"])])
]),
1000,
],
),
0,
],
)
aggregation_defs = {
"times_seen":
times_seen_aggregation,
"first_seen":
first_seen_aggregation,
"last_seen":
last_seen_aggregation,
# https://github.com/getsentry/sentry/blob/804c85100d0003cfdda91701911f21ed5f66f67c/src/sentry/event_manager.py#L241-L271
"priority":
Function(
"toUInt64",
[
Function(
"plus",
[
Function(
"multiply",
[
Function(
"log",
[times_seen_aggregation],
),
600,
],
),
last_seen_aggregation,
],
)
],
),
"user_count":
Function("ifNull", [
Function("uniq", [Column("tags[sentry:user]", entities["event"])]),
0
]),
}
def calculate_start_end(
self,
retention_window_start: Optional[datetime],
search_filters: Sequence[SearchFilter],
date_from: Optional[datetime],
date_to: Optional[datetime],
):
now = timezone.now()
end = None
end_params = [
_f for _f in
[date_to, get_search_filter(search_filters, "date", "<")] if _f
]
if end_params:
end = min(end_params)
if not end:
end = now + ALLOWED_FUTURE_DELTA
retention_date = max(
_f for _f in [retention_window_start, now - timedelta(days=90)]
if _f)
start_params = [
date_from, retention_date,
get_search_filter(search_filters, "date", ">")
]
start = max(_f for _f in start_params if _f)
end = max([retention_date, end])
return start, end, retention_date
def query(
self,
projects: Sequence[Project],
retention_window_start: Optional[datetime],
group_queryset: QuerySet,
environments: Sequence[Environment],
sort_by: str,
limit: int,
cursor: Optional[Cursor],
count_hits: bool,
paginator_options: Mapping[str, Any],
search_filters: Sequence[SearchFilter],
date_from: Optional[datetime],
date_to: Optional[datetime],
max_hits=None,
) -> CursorResult:
if not validate_cdc_search_filters(search_filters):
raise InvalidQueryForExecutor(
"Search filters invalid for this query executor")
start, end, retention_date = self.calculate_start_end(
retention_window_start, search_filters, date_from, date_to)
if start == retention_date and end == retention_date:
# Both `start` and `end` must have been trimmed to `retention_date`,
# so this entire search was against a time range that is outside of
# retention. We'll return empty results to maintain backwards compatibility
# with Django search (for now).
return self.empty_result
if start >= end:
# TODO: This maintains backwards compatibility with Django search, but
# in the future we should find a way to notify the user that their search
# is invalid.
return self.empty_result
e_event = self.entities["event"]
e_group = self.entities["group"]
where_conditions = [
Condition(Column("project_id", e_event), Op.IN,
[p.id for p in projects]),
Condition(Column("timestamp", e_event), Op.GTE, start),
Condition(Column("timestamp", e_event), Op.LT, end),
]
# TODO: This is still basically only handling status, handle this better once we introduce
# more conditions.
for search_filter in search_filters:
where_conditions.append(
Condition(Column(search_filter.key.name, e_group), Op.IN,
search_filter.value.raw_value))
if environments:
# TODO: Should this be handled via filter_keys, once we have a snql compatible version?
where_conditions.append(
Condition(Column("environment", e_event), Op.IN,
[e.name for e in environments]))
sort_func = self.aggregation_defs[self.sort_strategies[sort_by]]
having = []
if cursor is not None:
op = Op.GTE if cursor.is_prev else Op.LTE
having.append(Condition(sort_func, op, cursor.value))
query = Query(
"events",
match=Join([Relationship(e_event, "grouped", e_group)]),
select=[
Column("id", e_group),
replace(sort_func, alias="score"),
],
where=where_conditions,
groupby=[Column("id", e_group)],
having=having,
orderby=[OrderBy(sort_func, direction=Direction.DESC)],
limit=Limit(limit + 1),
)
data = snuba.raw_snql_query(
query, referrer="search.snuba.cdc_search.query")["data"]
hits_query = Query(
"events",
match=Join([Relationship(e_event, "grouped", e_group)]),
select=[
Function("uniq", [Column("id", e_group)], alias="count"),
],
where=where_conditions,
)
hits = None
if count_hits:
hits = snuba.raw_snql_query(
hits_query,
referrer="search.snuba.cdc_search.hits")["data"][0]["count"]
paginator_results = SequencePaginator(
[(row["score"], row["g.id"]) for row in data],
reverse=True,
**paginator_options,
).get_result(limit, cursor, known_hits=hits, max_hits=max_hits)
# We filter against `group_queryset` here so that we recheck all conditions in Postgres.
# Since replay between Postgres and Clickhouse can happen, we might get back results that
# have changed state in Postgres. By rechecking them we guarantee than any returned results
# have the correct state.
# TODO: This can result in us returning less than a full page of results, but shouldn't
# affect cursors. If we want to, we can iterate and query snuba until we manage to get a
# full page. In practice, this will likely only skip a couple of results at worst, and
# probably not be noticeable to the user, so holding off for now to reduce complexity.
groups = group_queryset.in_bulk(paginator_results.results)
paginator_results.results = [
groups[k] for k in paginator_results.results if k in groups
]
return paginator_results
def _render_trees(group: Group, user):
materialized_hashes = list({
gh.hash
for gh in GroupHash.objects.filter(project=group.project, group=group)
})
# Evaluates to the index of the last hash that is in materialized_hashes,
# or 1 otherwise.
find_hash_expr = _construct_arraymax([1] + [ # type: ignore
Function("indexOf", [Column("hierarchical_hashes"), hash])
for hash in materialized_hashes
])
# After much deliberation I (markus) decided that it would be best to
# render the entire tree using one large Snuba query. A previous
# implementation incurred n+1 queries on Snuba (n = number of materialized
# hashes) and was very buggy when it came to missing materialized hashes
# (which can happen if fallback/secondary grouping is turned on), events
# were counted twice because those n+1 queries accidentally counted
# overlapping sets of events, and the endpoint response time was kind of
# bad because of n+1 query.
#
# It being one large query may also make it easier to add pagination down
# the road.
query = (
Query("events", Entity("events")).set_select([
Function("count", [], "event_count"),
Function(
"argMax",
[Column("event_id"), Column("timestamp")], "event_id"),
Function("max", [Column("timestamp")], "latest_event_timestamp"),
# If hierarchical_hashes contains any of the materialized
# hashes, find_hash_expr evaluates to the last found index and
# arraySlice will give us this hash + the next child hash that
# we use in groupby
#
# If hierarchical_hashes does not contain any of those hashes,
# find_hash_expr will return 1 so we start slicing at the beginning.
# This can happen when hierarchical_hashes is empty (=>
# hash_slice = []), but we also try to recover gracefully from
# a hypothetical case where we are missing some hashes in
# postgres (unclear how this could be reached).
#
# We select some intermediate computation values here which we
# definitely don't need the results of. It's just temp vars.
Function(
# First we find the materialized hash using find_hash_expr,
# and subtract 1 which should be the parent hash if there
# is one. If there isn't, this now can be an out-of-bounds
# access by being 0 (arrays are indexed starting with 1)
"minus",
[find_hash_expr, 1],
"parent_hash_i",
),
# We clip the value to be at least 1, this will be where we
# start slicing hierarchical_hashes. 0 would be an out of
# bounds access.
Function("greatest", [Column("parent_hash_i"), 1], "slice_start"),
# This will return a slice of length 2 if the materialized hash
# has been found at the beginning of the array, but return a
# slice of length 3 if not.
Function(
"arraySlice",
[
Column("hierarchical_hashes"),
Column("slice_start"),
Function(
"minus",
[
Function(
"plus",
[Column("parent_hash_i"), 3],
),
Column("slice_start"),
],
),
],
"hash_slice",
),
Column("primary_hash"),
]).set_where(_get_group_filters(group)).set_groupby([
Column("parent_hash_i"),
Column("slice_start"),
Column("hash_slice"),
Column("primary_hash"),
]).set_orderby(
[OrderBy(Column("latest_event_timestamp"), Direction.DESC)]))
rv = []
for row in snuba.raw_snql_query(
query, referrer="api.group_split.render_grouping_tree")["data"]:
if len(row["hash_slice"]) == 0:
hash = row["primary_hash"]
parent_hash = child_hash = None
elif len(row["hash_slice"]) == 1:
(hash, ) = row["hash_slice"]
parent_hash = child_hash = None
elif len(row["hash_slice"]) == 2:
hash, child_hash = row["hash_slice"]
parent_hash = None
elif len(row["hash_slice"]) == 3:
parent_hash, hash, child_hash = row["hash_slice"]
else:
raise ValueError("unexpected length of hash_slice")
_add_hash(
rv,
group,
user,
parent_hash,
hash,
child_hash,
row["event_count"],
row["latest_event_timestamp"],
row["event_id"],
)
rv.sort(key=lambda tree: (tree["id"] or "", tree["childId"] or ""))
return rv
def _query_snuba(group: Group, id: int, offset=None, limit=None):
query = (Query("events", Entity("events")).set_select([
Function(
"arrayElement",
[
Column("hierarchical_hashes"),
Function("least", [
id + 1,
Function("length", [Column("hierarchical_hashes")])
]),
],
"new_materialized_hash",
),
Function("argMax",
[Column("event_id"), Column("timestamp")], "latest_event_id"),
Function("max", [Column("timestamp")], "latest_event_timestamp"),
Function("count", [], "event_count"),
]).set_groupby([Column("new_materialized_hash")]).set_orderby(
[OrderBy(Column("latest_event_timestamp"), Direction.DESC)]))
levels_overview = get_levels_overview(group)
# These conditions are always valid
common_where = [
Condition(Column("primary_hash"), Op.EQ,
levels_overview.only_primary_hash),
Condition(Column("project_id"), Op.EQ, group.project_id),
]
if id >= levels_overview.current_level:
# Good path: Since we increase the level we can easily constrain the
# entire query by group_id and timerange
query = query.set_where(common_where + _get_group_filters(group))
else:
# Bad path: We decreased the level and now we need to count events from
# other groups. If we cannot filter by group_id, we can also not
# restrict the timerange to anything at all. The Snuba API still
# requires us to set a timerange, so we set it to the maximum of 90d.
#
# Luckily the minmax index on group_id alone is reasonably efficient so
# that filtering by timerange (=primary key) is only a little bit
# faster.
now = datetime.datetime.now()
new_materialized_hash = _get_hash_for_parent_level(
group, id, levels_overview)
query = query.set_where(common_where + [
Condition(
Function("arrayElement",
[Column("hierarchical_hashes"), id + 1]),
Op.EQ,
new_materialized_hash,
),
Condition(Column("timestamp"), Op.GTE, now -
datetime.timedelta(days=90)),
Condition(Column("timestamp"), Op.LT, now +
datetime.timedelta(seconds=10)),
])
if offset is not None:
query = query.set_offset(offset)
if limit is not None:
query = query.set_limit(limit)
return snuba.raw_snql_query(
query, referrer="api.group_hashes_levels.get_level_new_issues")["data"]
def _render_trees(group: Group, user):
materialized_hashes = {
gh.hash
for gh in GroupHash.objects.filter(project=group.project, group=group)
}
rv = []
common_where = _get_group_filters(group)
for materialized_hash in materialized_hashes:
# For every materialized hash we want to render parent and child
# hashes, a limited view of the entire tree. We fetch one sample event
# so we know how we need to slice hierarchical_hashes.
hierarchical_hashes = _get_full_hierarchical_hashes(
group, materialized_hash)
if not hierarchical_hashes:
# No hierarchical_hashes found, the materialized hash is probably
# from flat grouping.
parent_pos = None
hash_pos = None
child_pos = None
slice_start = 0
else:
materialized_pos = hierarchical_hashes.index(materialized_hash)
if materialized_pos == 0:
parent_pos = None
hash_pos = 0
child_pos = 1
slice_start = 1
else:
parent_pos = 0
hash_pos = 1
child_pos = 2
slice_start = materialized_pos
# Select sub-views of the trees that contain materialized_hash.
query = (Query("events", Entity("events")).set_select([
Function("count", [], "event_count"),
Function(
"argMax",
[Column("event_id"), Column("timestamp")], "event_id"),
Function("max", [Column("timestamp")], "latest_event_timestamp"),
Function("arraySlice",
[Column("hierarchical_hashes"), slice_start, 3],
"hashes"),
]).set_where(common_where + [
Condition(
Function(
"has",
[
Column("hierarchical_hashes"),
materialized_hash,
],
),
Op.EQ,
1,
),
]).set_groupby([Column("hashes")]).set_orderby(
[OrderBy(Column("latest_event_timestamp"), Direction.DESC)]))
for row in snuba.raw_snql_query(query)["data"]:
assert not row["hashes"] or row["hashes"][
hash_pos] == materialized_hash
event_id = row["event_id"]
event = eventstore.get_event_by_id(group.project_id, event_id)
tree = {
"parentId": _get_checked(row["hashes"], parent_pos),
"id": materialized_hash,
"childId": _get_checked(row["hashes"], child_pos),
"eventCount": row["event_count"],
"latestEvent": serialize(event, user, EventSerializer()),
}
rv.append(tree)
if not row["hashes"]:
continue
try:
for variant in event.get_grouping_variants().values():
if not isinstance(variant, ComponentVariant):
continue
if variant.get_hash() == tree["parentId"]:
tree["parentLabel"] = variant.component.tree_label
if variant.get_hash() == tree["childId"]:
tree["childLabel"] = variant.component.tree_label
if variant.get_hash() == tree["id"]:
tree["label"] = variant.component.tree_label
except Exception:
sentry_sdk.capture_exception()
rv.sort(key=lambda tree:
(tree["parentId"] or "", tree["id"] or "", tree["childId"] or ""))
return rv
