Observability

One metrics runtime for region TPS, real tick durations, profiles, HUD bars, Spark, and stall traces.

Observability is not a separate dashboard bolted onto the server. Commands, HUD bars, profiles, and stall traces read the same samples. If two surfaces disagree, that is a bug, not a difference of opinion.

One runtime

Aurora measures region work in nanoseconds and reports durations in real time. A global TPS number is not fabricated to hide a region that is late. The runtime exposes the region view because regions are the execution model.

Commands

operator surface · illustrative output

> tps
  regions   20.0 / 20.0 / 19.9
  scheduler  EDF · 4 regions

> mspt
  p50 1.8 ms   p95 4.1 ms   max 12.6 ms

> perf start 30s
  sampling 30s…
  > perf stop
  region A  owner region-thread-a  ticks 600
  region B  owner region-thread-b  ticks 598

> aurora status
  engine    Aurora
  policy    EDF
  bridge    off

Command	Answers
/tps	Region tick rates and the scheduler view.
/mspt	Tick duration distribution with real nanosecond-derived values.
/perf	A bounded profile window; start and stop explicitly.
/aurora	Engine, scheduler policy, region count, bridge state.
/spark	Spark profiler integration when Spark is installed.
/sys	Cgroup-aware system view where the runtime can read it.

Illustrative outputThe values above are an illustration of the shape of the output, not a benchmark from a real workload. Do not use this page as a performance claim.

Stalls

The stall recorder captures the tick, the region, the owner thread, and the work that overran before the watchdog acts. A stall is a runtime event, not only a log line. The record should be enough to identify the owner and the operation that missed its deadline.

Build 48.4 also stops a stopped BRIDGE_IO lane from logging a warning and stack trace for every skipped repeating task during shutdown. A stopped lane says it is stopping, keeps its per-plugin rate limit, and stays quiet.

Reading the numbers

TPS is per region. A healthy global average can still contain a late region.
MSPT is a distribution. p50 alone can hide the tail that players feel.
Profiles are samples. Start and stop a bounded window; do not leave profiling on and call it production telemetry.
No claim without a workload. This release does not publish a benchmark or a player-count guarantee.

Continue

See Operations for commands in context, andConfiguration for the files that enable instrumentation.