Heron v0.8.3

[0.8.3] — 2026-10-12

Fixed

• Active calls climbed without bound and never came back down. A
  deployment whose real concurrency peaked at 28 reported a near-constant
  16,600 globally and 48,878 for one model. Concurrency was tracked as +1 per
  Start and −1 per Complete by dimension, and the two events name the model
  from different places: the Start from the request, the Complete from the
  response. A server that answers a request for one name with another — an
  alias, a canonical id — left every such call counted forever, and because
  the metrics shard is picked by a hash that includes the model, the Start and
  the Complete could also land on different shards, leaking the global count
  too. In-flight calls are now tracked by the call's own identity (connection
  4-tuple and request time, identical on both events) in one table shared by
  every shard, so the Complete closes exactly what its Start opened, and the
  global count is no longer split per shard. A Start whose Complete never
  arrives is closed after an hour of event time instead of lingering.
  New internal metrics: active_calls_open (calls in flight right now, which
  should track real concurrency) and active_calls_expired; they replace
  aggregator_concurrency_table. Metric rows already written keep their
  inflated values — only windows after the upgrade are correct.
• Filtering agent turns by model returned almost nothing on the aglake
  backend. aglake names an array field with a {} suffix at search time
  (models_used is searched as models_used{}, as Splunk does), and a search
  term on the bare name is not an error — it matches nothing. So the traces
  list filtered by model, and the agent-kind dropdown narrowed by model, came
  back empty or nearly so: one turn in seven days on a deployment where the
  model appeared in hundreds. The filters now use the suffixed name, and a live
  test filters by model — every earlier trace test passed an empty filter,
  which is how this got through.
• heron aglake-props listed models_used under its bare name, which
  aglake does not extract, so the field never reached the columnar store. It
  now prints models_used{}. Deployments that merged the earlier output
  should change that one line in props.toml and restart aglaked; as with any
  indexed change it applies to newly written buckets only.

Changed

• Performance page: token throughput instead of token totals. "Cache
  Token Usage" and "Token Averages" are replaced by Prefill (Input) Tokens
  / s and Decode (Output) Tokens / s, across all calls. The old charts
  plotted per-bucket totals, whose height changed with the zoom level (a 5m
  bucket is five 1m buckets stacked); a rate reads the same at any
  granularity. The prefill chart draws cache reads and writes beside input
  rather than stacked under it, because whether a server's input count
  already includes cache reads varies — Anthropic's API excludes them,
  OpenAI's includes them. The bucket still in progress is left off, since its
  total is incomplete and would chart as a drop at the live edge. Per-stream
  generation speed remains the TPOT chart and the Overview's "Avg TPS".

What's Changed

• fix(metrics): track in-flight calls by identity, not by dimension by @vaderyang in #205
• fix(aglake): search array fields by their {} name, so the model filter matches by @vaderyang in #204
• feat(console): performance page charts prefill and decode tokens per second by @vaderyang in #206
• release: v0.8.3 — active calls that stay put, a model filter that matches, token rate charts by @vaderyang in #207

Full Changelog: v0.8.2...v0.8.3