Monthly•
2026-08-31

Native Histogram Queries & an Experimental Service Dependency Graph | Greptime Monthly Report - No. 87

Prometheus native histograms are queryable end to end, from PromQL aggregations to Prometheus HTTP responses. OTLP traces now yield a service dependency graph you can query with SQL, derived at read time and still experimental. Cumulative `OTLP` exponential histograms can be ingested with an opt-in option.
Native Histogram Queries & an Experimental Service Dependency Graph | Greptime Monthly Report - No. 87
On this page
Lei Huang
Software Engineer

Summary ​

August 2026 · 2026-08-01 – 2026-08-31

What landed this month:

  • Prometheus native histograms are queryable end to end, from PromQL aggregations to Prometheus HTTP responses.
  • OTLP traces now yield a service dependency graph you can query with SQL, derived at read time and still experimental.
  • Cumulative OTLP exponential histograms can be ingested with an opt-in option.

GreptimeDB shipped v1.2.0-beta.2 on August 21, the second beta of the v1.2 line: one-way skip_wal changes with ALTER TABLE, JSON2 extension and layout preparation, procedure and admin event observability, and Flow runtime observability. Two breaking changes are worth checking before upgrading. Soft-drop and table recovery are enterprise-only again (db#8747) — beta1 exposed them in OSS because the enterprise gate was missing. An OSS beta2 metasrv rejects gc.experimental_soft_drop.enable = true at startup and cannot recover or purge tables soft-dropped in beta1, so recover what you need first. Persisted native histogram fields also changed signedness (db#8824), with no migration, downgrade, or mixed-version path.

Contributors ​

Over the past month, 20 contributors merged a total of 166 PRs. Among them, 8 community contributors contributed 10 PRs. Welcome to our new contributors: @xhwhis, @tian1220A, @dhruvxvaishnav, @wy471x, @fzlzjerry, @grezzko!

Thanks to the community contributors who shipped this month:

Highlights ​

db#8850 feat(servers): expose native histograms over Prometheus HTTP ​

Prometheus native histograms are now usable end to end: write them through remote write v2 or OTLP, aggregate them with PromQL, and read them back over the Prometheus HTTP API, including /metadata, label discovery and query annotations. Compared with classic le buckets, you get high-resolution latency distributions without one series per bucket, and the cutover can be gradual: db#8874 evaluates histogram_quantile and histogram_fraction over classic, native and mixed inputs in a single scan.

This closed the six-PR series July's report flagged as in progress: sample-kind semantics (db#8758), ingest validation (db#8775), ranges that mix float and histogram samples (db#8784), vector operators (db#8798), and histogram-aware sum, avg, count, group and count_values (db#8848). If you already store native histograms, two things changed: the persisted fields' signedness (db#8824), and the config location, now under prom_store (db#8744).

db#8614 feat: read-time entity relationships graph over telemetry ​

Once OTLP traces are in GreptimeDB, the service dependency graph does not have to be built separately. Query greptime_private.semantic_entities and greptime_private.semantic_relationships with SQL to see which service calls which, with RED metrics on every edge. Derivation runs at read time and as the caller: each source table is authorized separately, cancellation and deadlines follow the caller, and the window comes from the observed_at filter in the query, defaulting to the last hour (db#8794). Dependencies that telemetry cannot show can be inserted by hand into greptime_private.semantic_relationships_declared.

It is still experimental: the tables sit in greptime_private, they are not documented, and the output schema may change. Services that only emit metrics need otlp.experimental_enable_resource_info, off by default. The rest of the month expanded what the graph recognizes: more derived edge types and built-in naming conventions for Prometheus, Kubernetes and gen_ai attributes (db#8836, db#8854, db#8880, db#8797, db#8904, db#8936).

db#8900 feat(otlp): support cumulative exponential histograms ​

If your OTel pipeline emits exponential histograms, you no longer have to convert them to fixed buckets before writing. Turn on otlp.experimental_enable_exponential_histogram, off by default, and cumulative ExponentialHistogram metrics are stored as ordinary native histograms: the same storage format, and no protocol-specific query path.

Cumulative temporality only. Scales -4..=8 map directly and finer resolutions are downscaled to 8; delta temporality, min/max, exemplars and OTel Arrow are not covered yet. Points that cannot be ingested are now reported rather than silently accepted, and a request mixing valid and invalid points returns partial success.

db#8768 feat: add admin function to discard unflushed data ​

As a last-resort recovery step, you can now discard a region's unflushed in-memory data without losing what is already on disk. ADMIN discard_unflushed_data(...) takes a region ID, or a table name to cover every region of that table, and leaves all persisted SST files in place.

On the engine side (db#8600), the WAL replay frontier is advanced and persisted before the memtables are swapped, so repeated calls are safe and a stale in-flight flush cannot republish discarded rows.

db#8734 feat(event): add event context to procedure events ​

greptime_private.events now records where a DDL came from. Each submitted procedure event carries a structured EventContext as JSONB that tells SQL, gRPC, Prometheus and OTLP apart, along with automatic schema-evolution paths (auto_create, auto_alter). Whether a table was altered by a person or by ingest is now a SQL query.

With the actor behind each event and admin function executions recorded as well (db#8849, db#8835, db#8856, db#8834), the events table is usable as an audit trail.

Also this month ​

  • Authorization: Bearer <token> now works on the /v1/ HTTP interfaces. The server treats the token as opaque and leaves validation to the UserProvider, so JWT and OIDC policy stays outside the database (db#8719).
  • information_schema.flow_statistics and SHOW FLOW STATUS show how a Flow is actually running; in distributed Flow, start_time and uptime_seconds are still NULL (db#8392).
  • Smaller additions: uddsketch_rank (db#8929) and mergeable stddev_pop state functions (db#8972), greptime standalone start --daemon on Unix (db#8960), and riscv64 cross-builds with release artifacts (db#8820).
  • Engine-side work continued: two-phase metric series scans are now on by default for eligible sparse metric regions and fall back to the legacy scan elsewhere, with experimental_series_scan_v2 to turn them off (db#8826, db#8703, db#8788, db#8926); JSON2 gained per-column DDL settings, v2 layout primitives and v2 query support, though v2 is still not the default layout (db#8745, db#8833, db#8895, db#8901, db#8940, db#8964).

Ecosystem Updates ​

Grafana Plugin ​

2 releases this month (v3.0.1, v3.0.3).

  • Queries run through a Go backend against /v1/sql, so panel SQL, time macros included, can be used directly in Grafana alert rules.
  • A Use OTel toggle auto-fills GreptimeDB-style logs and traces columns.
  • v3.0.3 fixes authentication.

Existing SQL panels, Query Builder configurations and dashboards keep working. The release post covers the change in full.

Dashboard ​

4 releases this month (v0.13.14, v0.13.13, v0.13.12, v0.13.11).

  • Dashboards can be exported as a full snapshot.
  • A command palette makes navigation and actions keyboard-driven.
  • Virtual-list tables support column resizing, the compact view uses each column's natural width, and there is a new width option.
  • Query results can go full screen.
  • Perses is updated to 0.54, and the guided tour is removed.
  • Connections no longer break after changing hosts.

Kubernetes Operator ​

1 release this month (v0.6.1).

  • Audit logs are configurable through the CRDs, including frontend groups.
  • Service ports can be customized.
  • The ingress CRD field is removed, which is a breaking change.

MCP Server ​

1 release this month (v0.5.2).

  • mcp dependency is pinned to >=1.8.0,<2 to prevent ModuleNotFoundError on fresh installs.

Good First Issues ​

These two issues are a reasonable size for a first contribution to GreptimeDB. Comment on one to pick it up.

Issue#8951 Metric engine ignores ttl set on a logical table, so the OTLP metrics write hint has no effect ​

The metric engine accepts a TTL on a logical table and then ignores it, so an OTLP write hint like x-greptime-hints: ttl=180d silently does nothing. Either propagate the TTL to the physical table, or reject it at CREATE TABLE and ALTER TABLE time with a clear error.

Keywords: Table Engine, Write Protocols, SQL

Difficulty: Medium

Issue#8620 Support configuring region engine options through environment variables ​

region_engine options can only be set in TOML today. The equivalent environment variable, GREPTIMEDB_STANDALONE__REGION_ENGINE__MITO__GLOBAL_WRITE_BUFFER_REJECT_SIZE, fails at startup instead of being applied. Make the region engine list follow the same layered configuration mapping as the rest of the config.

Keywords: Configuration, Standalone, Datanode

Difficulty: Medium

Get Involved ​

Open an issue for bugs and feature requests, or drop into the community channel.

Stay in the loop

Join our community

Get the latest updates and discuss with other users.

GitHubYouTubeSlackLinkedin