Collect metrics from every inference cluster - #470

Merged

dennis-upbound merged 44 commits into

modelplaneai:mainfrom

dennis-upbound:dennis/metrics-impl

Oct 2, 2026

Commits on Oct 1, 2026

Add the MetricMapping and TelemetryDestination kinds
dennis-upbound
committed
5dca9dc
Report whether a mapping and a destination actually work
dennis-upbound
committed
f13387d
Regenerate schemas for the telemetry kinds
dennis-upbound
committed
7cea2bd
Add the telemetry guide
dennis-upbound
committed
6ed1904
Compose the OpenTelemetry collector

dennis-upbound
and
claude
committed
0946764
Narrow the telemetry API to what's in use

dennis-upbound
and
claude
committed
6eaf26f
Run a value rewrite in the context that can reach the value

dennis-upbound
and
claude
committed
8ada9b6
Convert framebuffer memory to the unit its name claims

dennis-upbound
and
claude
committed
8c75825
Describe the metrics instead of the collector's transform language

dennis-upbound
and
claude
committed
a4cc790
Hold a metric name to the characters a metric name can have

dennis-upbound
and
claude
committed
82a582e
Regenerate the schema lock after rebasing

dennis-upbound
and
claude
committed
de140e3
Type the parts of a sink that belong to Modelplane

dennis-upbound
and
claude
committed
9912a90
Let a sink address its destination the way its exporter does

dennis-upbound
and
claude
committed
8fc7fca
Scrape the endpoints these components actually publish

dennis-upbound
and
claude
committed
d878697
Label serving pods with what their metrics belong to

dennis-upbound
and
claude
committed
aa50066
Name the cluster a series belongs to after the one an operator named

dennis-upbound
and
claude
committed
2350058
Assert the fleet's telemetry in the local end-to-end test

dennis-upbound
and
claude
committed
4f01123
Scrape the GPU exporter the cluster came with

dennis-upbound
and
claude
committed
2d7265c
Keep a series' identity when the backend flattens it

dennis-upbound
and
claude
committed
9c006ec
Carry only the identity a series is attributed to

dennis-upbound
and
claude
committed
dced0f8
Tell a deployment's replicas apart

dennis-upbound
and
claude
committed
1481878
Fold several metrics into one, and take a histogram's count

dennis-upbound
and
claude
committed
5400e3d
Resolve a cluster gateway's name where kube-dns is the resolver

dennis-upbound
and
claude
committed
0075504
Set the required aggregation in the metric-mapping fixture

dennis-upbound
and
claude
committed
96b9c80
Regenerate the schema lock after rebasing

dennis-upbound
and
claude
committed
4a35660
Name the processor that lifts the identity for what it does

dennis-upbound
and
claude
committed
2397158
Keep every series unique to whatever produced it

dennis-upbound
and
claude
committed
505835a
Say what a series actually carries

dennis-upbound
and
claude
committed
47bf569
Let a component say it counts in percent

dennis-upbound
and
claude
committed
313006e

Commits on Oct 2, 2026

Compile a mapping's unit conversion to scale_metric

dennis-upbound
and
claude
committed
f93cbcd
Scrape the decode engine and the endpoint picker

dennis-upbound
and
claude
committed
ef3313e
Stop the collector gating the serving stack's readiness

dennis-upbound
and
claude
committed
9a368aa
Copy a sink's credential to the clusters that mount it

dennis-upbound
and
claude
committed
ea13a08
Build the collector's config from every TelemetryDestination

dennis-upbound
and
claude
committed
a30b21a
Stop the composed Endpoints being mirrored into a second slice

dennis-upbound
and
claude
committed
9e6c9c7
Correct what the telemetry APIs and guide claim

dennis-upbound
and
claude
committed
8d00dc8
Map only the SGLang metrics SGLang actually publishes

dennis-upbound
and
claude
committed
96b22dc
Skip a telemetry object the current schema rejects

dennis-upbound
and
claude
committed
0fcb57d
Leans down guide, clarifying some sections, adds examples
tr0njavolta
committed
2b7ce70
Merge branch 'dennis/metrics-impl' into metrics-edits-docs
tr0njavolta
authored
f0355d0
Merge pull request #1 from tr0njavolta/metrics-edits-docs
dennis-upbound
authored
313bc9c
Correct what the telemetry guide claims after the edit pass

dennis-upbound
and
claude
committed
56c8040
Answer the rest of the guide's review comments

dennis-upbound
and
claude
committed
cc09442
Drop acrossReplicas from MetricMapping

dennis-upbound
and
claude
committed
191b0bf