v1.3.0-beta.2

Release date: October 15, 2026

GreptimeDB v1.3.0-beta.2 is the second beta of the 1.3 release and covers changes since v1.3.0-beta.1. Most changes are bug fixes, including fixes for wrong query results and for writes that could be lost or hidden.

👍 Highlights

Distributed write retries enabled by default

In distributed clusters, the frontend retries failed writes to datanodes. Writes that hit a region during migration, failover, or repartition are retried after a route refresh. Busy regions and Unavailable, Cancelled, or DeadlineExceeded errors are retried with exponential backoff and jitter. Regions that acknowledged success aren't retried. Retries cover inserts, deletes, bulk inserts, and logical metric tables; standalone mode doesn't retry.

Retries are on by default with these settings. Set max_retries = 0 to disable them:

[write_retry]
max_retries = 3
initial_backoff = "1500ms"
max_backoff = "10s"
backoff_multiplier = 2.0

See the Frontend configuration, #9542, #9543, and #9574.

Signals Explore in the bundled dashboard

Signals Explore is a drilldown view for metrics, logs, and traces, with label filters. Logs link to their traces, and traces link to their logs. Native and gauge histograms aren't charted yet. See dashboard#651.

Fixes for silently wrong query results

The following queries returned wrong results without an error:

• PromQL comparisons, min/max, topk/bottomk, sqrt, and round follow Prometheus float semantics. Previously, errors / requests > 0.05 was true for a NaN ratio, so alerts fired when there was no traffic (#9479).
• Binary operations on folded histograms, such as histogram_fraction(...) + on(host, shard) ..., could drop matching series at query parallelism above 1 (#9509).
• ORDER BY and LIMIT on logical metric tables could return misordered rows or repeated partial groups (#9522). Ordering over TQL scalar() results is also fixed (#9523).
• A subquery without a step uses 60 seconds, the Prometheus default. Previously, count_over_time(x[5m:]) in an instant query returned 300 instead of 5 (#9481).

Other PromQL fixes are listed under Bug Fixes.

Fulltext bloom index no longer skips matching rows

matches_term on a bloom fulltext index could return fewer rows than a full scan: matches_term(msg, 'world') missed hello_world, and error missed 错误error日志. The index now splits text where matches_term does: _ separates tokens, and Han text is stored as bigrams.

Case-sensitive LIKE can now use the index for pruning:

SELECT * FROM logs WHERE message LIKE '%disk quota exceeded%';

On a case-insensitive index, lower(message) LIKE ... and matches_term(lower(message), ...) also use it; ILIKE doesn't. Existing SSTs need an index rebuild, as described in Upgrade notes. See #9470.

Storage and routing fixes

• Bulk memtable merges and flushes could let an older value override a newer write or delete. Metric engine physical tables and flat-format tables use the bulk memtable (#9517).
• Closing a region flushes skip-WAL writes, so they survive a graceful datanode shutdown (#9440).
• Writes during repartition could go to the wrong region and be missing from query results (#9458).

Dashboard

The bundled dashboard moves from v0.13.15 to v0.14.2. Besides Signals Explore, it applies the selected timezone to both display and SQL requests (dashboard#657), and fixes log scrolling, table loading, and trace-ID lookups (dashboard#655, dashboard#656, dashboard#658).

Benchmark results

Compared with v1.2.1 on the same setup for both versions. The results cover all changes since v1.2.1, not only this beta.

• 30-day PromQL range queries over node metrics (1,720 series, 158M samples, 30m step, warm cache): the five queries are 1.14×–1.83× faster at the median (geometric mean 1.43×). Throughput is 16% higher at concurrency 1 and 6% higher at concurrency 4 (geometric means); one query is 8% slower at concurrency 4.
• SQL queries over 100M agent observability records (c7i.2xlarge, 8 CPUs and 16 GB for the database): the 16 queries are 1.08×–1.44× faster at the median (geometric mean 1.20×).

Upgrade notes

• Distributed PromQL: Upgrade datanodes before frontends. Older datanodes can't decode the new PromQL functions in distributed query plans (#9479).
• Fulltext bloom indexes: matches_term and LIKE scan SSTs with old-format indexes until compaction or ADMIN BUILD_INDEX('logs') rebuilds them. To check rebuild progress, read fulltext.bloom_version in the meta_json column of information_schema.ssts_index_meta. After a downgrade, older binaries also scan SSTs with new-format indexes. The bloom backend no longer uses jieba; Han text is stored as bigrams, which can make the index larger. Tantivy-backed indexes no longer serve matches_term; those queries scan (#9470).
• Write retries: If every attempt fails, the error returns after at least 10.5 seconds of backoff (1.5 + 3 + 6) plus jitter. Clients with shorter timeouts see a timeout instead. If a write succeeds but its response is lost or times out, the retry can insert duplicate rows into append-only tables. Set max_retries = 0 under [write_retry] to disable retries (#9543).
• Flow mirroring: When a flownode is slow or unreachable and more than 1M mirrored rows are pending, the frontend drops further mirrored batches to protect its memory. Flow results can have gaps (#9421).
• Series scans: The v2 series-scan path is temporarily disabled by default (#9436).

Breaking changes

• Expression depth: Query planning rejects expressions deeper than 128 levels to reduce stack-overflow risk (#9564).
• PromQL subquery step: Range queries that relied on [R:] using the query step now sample every 60 seconds. The step isn't configurable (#9481).
• Decompressed request size: Prometheus remote read/write and Loki push bodies over 512 MiB after decompression are rejected with 400. With a memory quota configured, compressed gRPC requests reserve max_recv_message_size before decompression, so 429 or RESOURCE_EXHAUSTED can occur sooner (#9264).

🚀 Features

• feat(json2): json_get respect type hints by @fengys1996 in #9349
• feat: update dashboard to v0.14.0 by @sunchanglong in #9510
• feat: support automatic column addition for Flight bulk inserts by @v0y4g3r in #9285
• feat(catalog): declare information-schema table scopes by @shuiyisong in #9504
• feat(meta-client): forward actor and event context for region follower RPCs by @sunng87 in #9568
• feat: update dashboard to v0.14.2 by @sunchanglong in #9570
• feat(client): support region request reuse and route refresh hints by @WenyXu in #9540
• feat(operator): retry failed writes with refreshed partitions by @WenyXu in #9542
• feat(frontend): retry transient distributed write failures by @WenyXu in #9543

🐛 Bug Fixes

• fix: temporarily disable v2 series scan by default by @evenyag in #9436
• fix: align TWCS assignment with half-open time windows by @v0y4g3r in #9464
• fix(json2): preserve metadata when adding columns via alter table by @fengys1996 in #8696
• fix: bound decompressed request size and close memory-admission gaps for compressed requests by @sunng87 in #9264
• fix(promql): apply offset to subquery evaluation window by @DeviousCardi in #9364
• fix(client): yield Flight batches and affected rows without waiting for next message by @discord9 in #8918
• fix(promql): resolve dotted column names as unqualified columns by @killme2008 in #9391
• fix(promql): align timestamp(), label_join and label_replace with Prometheus semantics by @killme2008 in #9385
• fix(mito2): truncate parquet column index min/max for SST writes by @killme2008 in #9420
• fix(catalog): skip backend batch_get on full cache hits by @WenyXu in #9431
• fix: coalesce equal histogram quantile bucket boundaries by @evenyag in #9451
• fix: accept a bool value for the CLI --auto-create-schema flag by @fengyue-xve in #9450
• fix(otlp): coerce empty and "-" trace attribute strings to null by @killme2008 in #9461
• fix(meta-client): bound heartbeat handshake timeout by @Aetherance in #9438
• fix(partition): preserve region mapping during repartition by @killme2008 in #9458
• fix(promql): restore exact filtering for nullable label matchers by @killme2008 in #9418
• fix(metasrv): retry failed heartbeat address updates by @WenyXu in #9437
• fix(operator): bound FlowMirrorTask pending rows to prevent frontend OOM by @discord9 in #9421
• fix(promql): align subqueries to the absolute evaluation grid by @discord9 in #9400
• fix(promql): keep vector labels when a computed scalar is on the right by @killme2008 in #9478
• fix(otlp): convert boolean metric attributes into label values by @fengyue-xve in #9468
• fix(promql): return no point for zero-interval irate/idelta by @niladrix719 in #9455
• fix(metasrv): stop parking the heartbeat response on the region supervisor queue by @lczllx in #9474
• fix(promql): keep metric name for comparison operators by @Aumnertic in #9332
• fix(metasrv): stop holding the pusher registry lock across mailbox sends by @lczllx in #9445
• fix(mito): flush memtables containing skip-WAL writes on region close by @WenyXu in #9440
• fix(promql): accept evaluation times before the Unix epoch by @killme2008 in #9480
• fix: unify unknown flow option rejection across OSS and enterprise builds by @discord9 in #9319
• fix(promql): return range-vector results for instant queries by @killme2008 in #9477
• fix(query): stop reporting region scans as hash partitioned by @discord9 in #9491
• fix(table): require complete primary keys for hash partitioning by @discord9 in #9507
• fix(deps): bump OpenDAL to 0.59.4 for OSS RRSA credentials by @WenyXu in #9520
• fix(promql): default a stepless subquery to a one-minute step by @killme2008 in #9481
• fix(promql): follow IEEE 754 float semantics in comparisons and aggregations by @killme2008 in #9479
• fix(query): keep filter fetch when rewriting matches term by @discord9 in #9521
• fix(mito): preserve version ordering in bulk merges and flushes by @evenyag in #9517
• fix(mito2): ignore missing inverted indexes in old SSTs by @WenyXu in #9545
• fix: prevent stale leader metadata after concurrent writes by @fengjiachun in #9549
• fix(promql): stop fabricating histogram fold hash partitioning by @discord9 in #9509
• fix(table): stop claiming tag order for logical metric scans by @discord9 in #9522
• fix(promql): stop claiming scalar calculate preserves order by @discord9 in #9523
• fix: align fulltext bloom tokens with matches_term and use them for LIKE by @killme2008 in #9470
• fix(query): guard deep expressions during query planning by @evenyag in #9564
• fix(operator): reuse catalog table resolution for write retries by @WenyXu in #9574
• fix(json2): support settings in ALTER TABLE ADD COLUMN by @fengys1996 in #9472
• fix(query): accept narrowed nullability in MergeScan remote schema check by @killme2008 in #9592
• fix(table): stop claiming exact null and distinct counts for append scans by @killme2008 in #9593
• fix: avoid pinning memtables during series-index reconciliation by @evenyag in #9588
• fix(query): keep state column names unique when splitting aggregates by @killme2008 in #9594
• build(deps): update opensrv-mysql to fix MariaDB 12.3 USE information_schema hang by @tian1220A in #9357

🚜 Refactor

• refactor: expose OTLP write and batching helpers for downstream routers by @shuiyisong in #9401
• refactor(operator): centralize write partitioning and dispatch by @WenyXu in #9541

📚 Documentation

• docs(config): align example configs and config.md with actual defaults by @killme2008 in #9460

⚡ Performance

• perf(catalog): optimize single-table lookups in information_schema.partitions by @fengyue-xve in #9473
• perf(promql): reduce timestamp payload allocations and use narrow cursor scans by @discord9 in #9417
• perf(query): infer safe aggregate input bounds from HAVING by @v0y4g3r in #9582

🧪 Testing

• test(sqlness): wait for mirrored inserts before flushing streaming flows by @killme2008 in #9424
• test(promql): cover label_replace normalization with distinct series by @fengyue-xve in #9485

⚙️ Miscellaneous Tasks

• ci(query-regression): bump RUNNER_IMAGE_EPOCH to 7 by @MichaelScofield in #9395
• ci: retry failed nightly release on following weekdays by @sunng87 in #9394
• ci: upgrade EC2 runner action to restore runner registration by @evenyag in #9430
• ci: make query regression non-blocking for scheduled nightly releases by @sunng87 in #9434
• chore: update rust toolchain to 1.98.1 by @sunng87 in #9429
• ci: update dev-builder image tag by @github-actions[bot] in #9469
• ci: enhance benchmark workflows with version comparisons and diagnostics by @WenyXu in #9383
• ci: benchmark the object store WAL against S3 from an EC2 runner by @fengjiachun in #9536
• build: reduce checkout-specific inputs for cache reuse by @discord9 in #9198

New Contributors

• @Aumnertic made their first contribution in #9332
• @Aetherance made their first contribution in #9438
• @fengyue-xve made their first contribution in #9450
• @niladrix719 made their first contribution in #9455

All Contributors

We would like to thank the following contributors from the GreptimeDB community:

@Aetherance, @Aumnertic, @DeviousCardi, @MichaelScofield, @WenyXu, @discord9, @evenyag, @fengjiachun, @fengys1996, @fengyue-xve, @killme2008, @lczllx, @niladrix719, @shuiyisong, @sunchanglong, @sunng87, @tian1220A, @v0y4g3r