Skip to content

known-limitations.md

Latest commit

 

History

History
37 lines (21 loc) · 4.16 KB

known-limitations.md

File metadata and controls

37 lines (21 loc) · 4.16 KB
Raw

Known limitations

Openfuse is a beta. The GreptimeDB migration is functionally complete and parity-verified for the covered surface. This page lists the genuine limitations — constraints and caveats that affect how you deploy, query, or recover.

It is not a list of output differences. Where the fork's dashboard/metrics output diverges from upstream Langfuse, the fork is equal or more correct; those intentional differences are summarised at the end and detailed in the parity ledger.

Functional

Indexed full-text search is whole-term only

Full-text indexes on input, output, and metadata values back the explicit FTS operator through whole-term matches_term. Substring matching (contains / starts with / ends with) is not indexed — it compiles to case-insensitive LIKE and scans, which is correct but can be slow on large payload columns over wide time windows.

Operational

GreptimeDB migrations are idempotent DDL, with no ledger

The web and standalone entrypoints apply the GreptimeDB schema on startup (gated by LANGFUSE_AUTO_GREPTIME_MIGRATION_DISABLED, serialised across replicas by a Postgres advisory lock). There is no applied-migrations ledger, so the runner re-applies every .sql on each start and every statement must stay idempotent: CREATE ... IF NOT EXISTS, declarative ALTER ... SET INDEX, applied in file order, with no down-migrations. It tolerates only the one common non-idempotent re-run error (ADD COLUMN on an existing column, errno 1060); any other migration error fails the container at startup. A migration that rewrites existing data must be written with this in mind, and a real migration ledger is on the roadmap. See deployment.

Object storage is optional, not gone

Ingestion needs no object store — traces, observations, and scores persist to GreptimeDB raw_events. Media uploads, the OTel carrier, the eval blob store, and batch exports all default to local filesystem volumes in the bundled Compose files, so a stock deployment runs with no S3/MinIO at all. Object storage stays fully supported (set the *_STORAGE_BACKEND variables to s3) for multi-node deployments, where a shared object store is the simpler way to share blobs across replicas. See deployment.

Source-of-truth exception

dataset_run_items deletion is a projection hard-delete

Traces, observations, and scores delete by appending a tombstone to raw_events, so replay rebuilds a soft-deleted row and a restored raw_events reconstructs deletion state. dataset_run_items instead hard-deletes the projection and relies on the replay path skipping deleted items; it keeps no per-entity tombstone in raw_events. So "raw_events is the complete source of truth" holds for traces/observations/scores but not for DRI. It works operationally; it is called out so the exception is explicit.

Intentional differences from upstream Langfuse (not limitations)

These are documented behaviour differences, not defects: output can differ from upstream, but the fork is equal or more correct. Full detail and parity evidence are in the parity ledger.

  • Approximate percentiles. p50–p99 use GreptimeDB uddsketch; upstream uses a different approximation. Values differ slightly; both are approximate.
  • Empty time buckets are filled. Time-bucketed queries emit a zero-valued row for every empty bucket in range; upstream omits them. Buckets with data match exactly.
  • Stricter query validation. The fork returns 400 InvalidRequestError for some nonsensical dashboard queries that upstream silently accepts (e.g. histogram on a relation-backed measure, sum on a string measure).
  • By-called-tool value breakdowns conserve. Breaking cost/tokens down by called tool attributes an observation's value once per distinct tool; upstream multiplies by call multiplicity. Tool filters and available-tool breakdowns are exact parity.
  • Larger default model-price catalog. The fork bundles more default model prices than the tracked upstream release; per-model cost is identical. Pin the same catalog for byte-identical GET /api/public/models.