prometheus: scrape per-track metrics from /metrics/track

Namespace is a required parameter and there is no enumeration endpoint, so
namespaces are targets relabelled into the query. The static list should be
generated from /state.namespace_tree into a file_sd file; noted in the README.

Needs #533.

Author: gmarzot

docker/prometheus/README.md

@@ -0,0 +1,31 @@
+# Prometheus scrape config
+
+## Per-track metrics (`moqx-track`)
+
+`/metrics/track` requires a `namespace` parameter, so every namespace to be
+graphed is a scrape target that relabelling turns into the query parameter.
+Namespace values use the moq-transport encoded form — tuple elements joined by
+`-`, other bytes as `.<hex>`:
+
+    moq-test/interop   ->  moq.2dtest-interop
+    conf.example.com / room 1  ->  conf.2eexample.2ecom-room.201
+
+`limit` is a guard rather than a selector: a namespace holding more tracks than
+the limit returns 400, so an over-wide scrape fails visibly instead of graphing
+a series set that reshuffles between scrapes. Ceiling is
+`admin.track_metrics_max_limit` (1000).
+
+### Generating the target list
+
+The static list needs maintaining by hand. `/state.namespace_tree` already
+enumerates live namespaces, so the intended replacement is a small generator
+that walks the tree, encodes each namespace, and writes a `file_sd` target
+file; Prometheus reloads target files without a restart.
+
+    file_sd_configs:
+      - files: ['/etc/prometheus/targets/namespaces.json']
+        refresh_interval: 30s
+
+Namespaces churn as events start and end, and a relay may hold more than the
+per-scrape ceiling, so this becomes necessary rather than convenient once the
+relay carries production traffic.

docker/prometheus/prometheus.yml

@@ -72,3 +72,34 @@ scrape_configs:
         target_label: __param_target
       - target_label: __address__
         replacement: json-exporter:7979
+
+  # Per-track counters from the relay (openmoq/moqx#533). The namespace
+  # parameter is required, so each namespace is a target relabelled into the
+  # query. Values use the moq-transport encoded form: tuple elements joined
+  # by "-", other bytes as ".<hex>" (moq-test/interop -> moq.2dtest-interop).
+  #
+  # limit is a guard, not a top-N: a namespace holding more tracks than the
+  # limit is rejected with 400 rather than truncated, so the scrape fails
+  # loudly instead of graphing a set that reshuffles between scrapes.
+  #
+  # The list is static for now. /state.namespace_tree already enumerates live
+  # namespaces, so it should be generated into a file_sd target file and
+  # picked up without a reload — see docker/prometheus/README.md.
+  - job_name: moqx-track
+    metrics_path: /metrics/track
+    scrape_interval: 15s
+    params:
+      limit: ['1000']
+    static_configs:
+      - targets:
+          - moq.2dtest-interop
+          - aiomoqt
+    relabel_configs:
+      - source_labels: [__address__]
+        target_label: __param_namespace
+      - source_labels: [__param_namespace]
+        target_label: ns_scrape
+      - target_label: __address__
+        replacement: moqx:8000
+      - target_label: instance
+        replacement: moqx-relay