feat(seer grouping): Store Seer metadata on grouphashes during ingest

fixtures/backup/model_dependencies/detailed.json

@@ -2319,6 +2319,16 @@
  "kind": "DefaultOneToOneField",
  "model": "sentry.grouphash",
  "nullable": false
+ },
+ "seer_grouphash_sent": {
+ "kind": "FlexibleForeignKey",
+ "model": "sentry.grouphash",
+ "nullable": true
+ },
+ "seer_matched_grouphash": {
+ "kind": "FlexibleForeignKey",
+ "model": "sentry.grouphash",
+ "nullable": true
  }
  },
  "model": "sentry.grouphashmetadata",

migrations_lockfile.txt

@@ -10,7 +10,7 @@ hybridcloud: 0016_add_control_cacheversion
 nodestore: 0002_nodestore_no_dictfield
 remote_subscriptions: 0003_drop_remote_subscription
 replays: 0004_index_together
-sentry: 0764_migrate_bad_status_substatus_rows
+sentry: 0765_add_seer_fields_to_grouphash_metadata
 social_auth: 0002_default_auto_field
 uptime: 0013_uptime_subscription_new_unique
 workflow_engine: 0005_data_source_detector

src/sentry/event_manager.py

@@ -1445,7 +1445,7 @@ def _save_aggregate(
  # If we still haven't found a matching grouphash, we're now safe to go ahead and talk to
  # seer and/or create the group.
  if existing_grouphash is None:
- seer_matched_grouphash = maybe_check_seer_for_matching_grouphash(event)
+ seer_matched_grouphash = maybe_check_seer_for_matching_grouphash(event, grouphashes)
  seer_matched_group = (
  Group.objects.filter(id=seer_matched_grouphash.group_id).first()
  if seer_matched_grouphash
@@ -1614,7 +1614,7 @@ def _save_aggregate_new(
  result = "found_secondary"
  # If we still haven't found a group, ask Seer for a match (if enabled for the project)
  else:
- seer_matched_grouphash = maybe_check_seer_for_matching_grouphash(event)
+ seer_matched_grouphash = maybe_check_seer_for_matching_grouphash(event, all_grouphashes)
 
  if seer_matched_grouphash:
  group_info = handle_existing_grouphash(job, seer_matched_grouphash, all_grouphashes)
@@ -1806,16 +1806,6 @@ def _create_group(
  first_release: Release | None = None,
  **group_creation_kwargs: Any,
 ) -> Group:
- # Temporary log to debug events seeming to disappear after being sent to Seer
- if event.data.get("seer_similarity"):
- logger.info(
- "seer.similarity.pre_create_group",
- extra={
- "event_id": event.event_id,
- "hash": event.get_primary_hash(),
- "project": project.id,
- },
- )
 
  short_id = _get_next_short_id(project)
 
@@ -1891,18 +1881,6 @@ def _create_group(
  logger.exception("Error after unsticking project counter")
  raise
 
- # Temporary log to debug events seeming to disappear after being sent to Seer
- if event.data.get("seer_similarity"):
- logger.info(
- "seer.similarity.post_create_group",
- extra={
- "event_id": event.event_id,
- "hash": event.get_primary_hash(),
- "project": project.id,
- "group_id": group.id,
- },
- )
-
  return group
 
 

src/sentry/grouping/ingest/seer.py

@@ -232,7 +232,9 @@ def get_seer_similar_issues(
  return (similar_issues_metadata, parent_grouphash)
 
 
-def maybe_check_seer_for_matching_grouphash(event: Event) -> GroupHash | None:
+def maybe_check_seer_for_matching_grouphash(
+ event: Event, all_grouphashes: list[GroupHash]
+) -> GroupHash | None:
  seer_matched_grouphash = None
 
  if should_call_seer_for_grouping(event):
@@ -241,16 +243,48 @@ def maybe_check_seer_for_matching_grouphash(event: Event) -> GroupHash | None:
  sample_rate=options.get("seer.similarity.metrics_sample_rate"),
  tags={"call_made": True, "blocker": "none"},
  )
+
  try:
  # If no matching group is found in Seer, we'll still get back result
  # metadata, but `seer_matched_grouphash` will be None
  seer_response_data, seer_matched_grouphash = get_seer_similar_issues(event)
- event.data["seer_similarity"] = seer_response_data
-
- # Insurance - in theory we shouldn't ever land here
- except Exception as e:
+ except Exception as e: # Insurance - in theory we shouldn't ever land here
  sentry_sdk.capture_exception(
  e, tags={"event": event.event_id, "project": event.project.id}
  )
+ return None
+
+ # Find the GroupHash for the hash value sent to Seer
+ primary_hash = event.get_primary_hash()
+ grouphash_sent = list(
+ filter(lambda grouphash: grouphash.hash == primary_hash, all_grouphashes)
+ )[0]
+
+ # Update GroupHashes with Seer results
+ for grouphash in all_grouphashes:
+ metadata = grouphash.metadata
+
+ if metadata:
+ # Mark all the GroupHashes as having been represented by the one we sent
+ metadata.seer_grouphash_sent = grouphash_sent
+
+ # Store the Seer results only on the GroupHash which was actually sent
+ if grouphash is grouphash_sent:
+ # Technically the time of the metadata record creation and the time of the Seer
+ # request will be some milliseconds apart, but the difference isn't meaningful
+ # and forcing them to be the same (rather than just close) lets us use their
+ # equality as a signal that the Seer call happened during ingest rather than
+ # during a backfill, without having to store that information separately
+ metadata.seer_date_sent = metadata.date_added
+ metadata.seer_event_sent = event.event_id
+ metadata.seer_model = seer_response_data["similarity_model_version"]
+ metadata.seer_matched_grouphash = seer_matched_grouphash
+ metadata.seer_match_distance = (
+ seer_response_data["results"][0]["stacktrace_distance"]
+ if seer_matched_grouphash
+ else None
+ )
+
+ metadata.save()
 
  return seer_matched_grouphash

src/sentry/migrations/0765_add_seer_fields_to_grouphash_metadata.py

@@ -0,0 +1,70 @@
+# Generated by Django 5.1.1 on 2024-09-23 15:40
+
+import django.db.models.deletion
+from django.db import migrations, models
+
+import sentry.db.models.fields.foreignkey
+from sentry.new_migrations.migrations import CheckedMigration
+
+
+class Migration(CheckedMigration):
+ # This flag is used to mark that a migration shouldn't be automatically run in production.
+ # This should only be used for operations where it's safe to run the migration after your
+ # code has deployed. So this should not be used for most operations that alter the schema
+ # of a table.
+ # Here are some things that make sense to mark as post deployment:
+ # - Large data migrations. Typically we want these to be run manually so that they can be
+ # monitored and not block the deploy for a long period of time while they run.
+ # - Adding indexes to large tables. Since this can take a long time, we'd generally prefer to
+ # run this outside deployments so that we don't block them. Note that while adding an index
+ # is a schema change, it's completely safe to run the operation after the code has deployed.
+ # Once deployed, run these manually via: https://develop.sentry.dev/database-migrations/#migration-deployment
+
+ is_post_deployment = False
+
+ dependencies = [
+ ("sentry", "0764_migrate_bad_status_substatus_rows"),
+ ]
+
+ operations = [
+ migrations.AddField(
+ model_name="grouphashmetadata",
+ name="seer_date_sent",
+ field=models.DateTimeField(null=True),
+ ),
+ migrations.AddField(
+ model_name="grouphashmetadata",
+ name="seer_event_sent",
+ field=models.CharField(max_length=32, null=True),
+ ),
+ migrations.AddField(
+ model_name="grouphashmetadata",
+ name="seer_grouphash_sent",
+ field=sentry.db.models.fields.foreignkey.FlexibleForeignKey(
+ null=True,
+ on_delete=django.db.models.deletion.DO_NOTHING,
+ related_name="+",
+ to="sentry.grouphash",
+ ),
+ ),
+ migrations.AddField(
+ model_name="grouphashmetadata",
+ name="seer_match_distance",
+ field=models.FloatField(null=True),
+ ),
+ migrations.AddField(
+ model_name="grouphashmetadata",
+ name="seer_matched_grouphash",
+ field=sentry.db.models.fields.foreignkey.FlexibleForeignKey(
+ null=True,
+ on_delete=django.db.models.deletion.DO_NOTHING,
+ related_name="seer_matchees",
+ to="sentry.grouphash",
+ ),
+ ),
+ migrations.AddField(
+ model_name="grouphashmetadata",
+ name="seer_model",
+ field=models.CharField(null=True),
+ ),
+ ]

src/sentry/models/grouphashmetadata.py

@@ -4,18 +4,54 @@
 from sentry.backup.scopes import RelocationScope
 from sentry.db.models import Model, region_silo_model
 from sentry.db.models.base import sane_repr
+from sentry.db.models.fields.foreignkey import FlexibleForeignKey
 
 
 @region_silo_model
 class GroupHashMetadata(Model):
  __relocation_scope__ = RelocationScope.Excluded
 
  # GENERAL
+
  grouphash = models.OneToOneField(
  "sentry.GroupHash", related_name="_metadata", on_delete=models.CASCADE
  )
  date_added = models.DateTimeField(default=timezone.now)
 
+ # SEER
+
+ # Only one hash representing each group is sent to Seer. For the grouphash actually sent, this
+ # field and the `grouphash` field will be identical. For the grouphashes assigned to the same
+ # group but which aren't sent, this will point to the GroupHash record for the sent hash. Note
+ # that because of merging/unmerging, the sent GroupHash and this metadata's GroupHash (if not
+ # one and the same) aren't guaranteed to forever point to the same group (though they will when
+ # this field is written).
+ seer_grouphash_sent = FlexibleForeignKey(
+ "sentry.GroupHash",
+ # If we end up needing to reference in this direction, we can handle it with a property on
+ # GroupHash
+ related_name="+",
+ on_delete=models.DO_NOTHING,
+ null=True,
+ )
+
+ # NOTE: The rest of the Seer-related fields are only stored on the metadata of the GroupHash
+ # actually sent to Seer.
+
+ # When this hash was sent to Seer. This will be different than `date_added` if we send it to
+ # Seer as part of a backfill rather than during ingest.
+ seer_date_sent = models.DateTimeField(null=True)
+ # Id of the event whose stacktrace was sent to Seer
+ seer_event_sent = models.CharField(max_length=32, null=True)
+ # The version of the Seer model used to process this hash value
+ seer_model = models.CharField(null=True)
+ # The `GroupHash` record representing the match Seer sent back as a match (if any)
+ seer_matched_grouphash = FlexibleForeignKey(
+ "sentry.GroupHash", related_name="seer_matchees", on_delete=models.DO_NOTHING, null=True
+ )
+ # The similarity between this hash's stacktrace and the parent (matched) hash's stacktrace
+ seer_match_distance = models.FloatField(null=True)
+
  class Meta:
  app_label = "sentry"
  db_table = "sentry_grouphashmetadata"

tests/sentry/backup/snapshots/test_comparators/test_default_comparators.pysnap

@@ -1,5 +1,5 @@
 ---
-created: '2024-09-17T12:19:51.014511+00:00'
+created: '2024-09-23T17:48:55.681597+00:00'
 creator: sentry
 source: tests/sentry/backup/test_comparators.py
 ---
@@ -579,6 +579,8 @@ source: tests/sentry/backup/test_comparators.py
  - class: ForeignKeyComparator
  fields:
  - grouphash
+ - seer_grouphash_sent
+ - seer_matched_grouphash
  model_name: sentry.grouphashmetadata
 - comparators:
  - class: ForeignKeyComparator