feat(metrics): 3 Prometheus counters for scaling dive (#7756)

[JohnMcLear]

Summary

Adds three Prometheus rows to `/stats/prometheus` so the load-test harness for #7756 can attribute where time goes on the server, not just the gauge headline (CPU / event-loop / memory):

• `etherpad_pad_users{padId}` — gauge derived from `sessioninfos` on scrape. Lets the harness confirm a target pad actually has the expected concurrency.
• `etherpad_changeset_apply_duration_seconds` — histogram observed inside `handleUserChanges`. Separates "apply path is slow" from "fan-out is slow" when latency rises.
• `etherpad_socket_emits_total{type}` — counter at the broadcast emit sites (`handleCustomObjectMessage`, `handleCustomMessage`, `sendChatMessageToPadClients`) and the per-socket `NEW_CHANGES` loop in `updatePadClients`. Bucketed by message type so the harness can measure the amplification factor of each lever (especially the fan-out batching lever).

Metric handles live in a new `src/node/prom-instruments.ts` rather than `prometheus.ts` itself, to avoid a circular import — `prometheus.ts` already `require`s `PadMessageHandler` to read `sessioninfos`.

This is the follow-up PR planned in section 6 of `docs/superpowers/specs/2026-05-15-scaling-dive-design.md` (in `ether/etherpad-load-test`). The load-test harness already tolerates these metrics' absence, so it ships unblocked; this PR just makes the next dive run more informative.

Test Plan

• Backend test `src/tests/backend-new/specs/prom-instruments.test.ts` passes (3/3 locally) — verifies the recording helpers move the underlying counter/histogram.
• Existing backend test suite remains green.
• After merge, `curl http://your-etherpad/stats/prometheus | grep etherpad_` shows the three new rows.
• Trigger the Scaling dive workflow with `core_ref` pointing at this branch or develop-post-merge; verify the harness reports include `etherpad_socket_emits_total{type=NEW_CHANGES}` and `etherpad_changeset_apply_duration_seconds_count` in the JSON snapshot.

🤖 Generated with Claude Code

[qodo-code-review]

Qodo reviews are paused for this user.

Troubleshooting steps vary by plan Learn more →

On a Teams plan?
Reviews resume once this user has a paid seat and their Git account is linked in Qodo.
Link Git account →

Using GitHub Enterprise Server, GitLab Self-Managed, or Bitbucket Data Center?
These require an Enterprise plan - Contact us
Contact us →

[qodo-free-for-open-source-projects]

Review Summary by Qodo

Add Prometheus metrics for scaling dive load testing

✨ Enhancement

Walkthroughs

Description

• Add three Prometheus metrics for scaling dive load testing
  - etherpad_pad_users{padId} gauge tracks concurrent users per pad
  - etherpad_changeset_apply_duration_seconds histogram measures edit processing time
  - etherpad_socket_emits_total{type} counter tracks broadcast messages by type
• Create new prom-instruments.ts module to avoid circular imports
• Instrument hot paths in PadMessageHandler to record metrics
• Add smoke tests verifying metric recording functionality

Diagram

flowchart LR
  PMH["PadMessageHandler<br/>hot path"] -->|import| PI["prom-instruments.ts<br/>metric handles"]
  PI -->|register| PROM["prometheus.ts<br/>central registry"]
  PMH -->|recordSocketEmit| SE["socketEmitsTotal<br/>counter"]
  PMH -->|recordChangesetApply| CAD["changesetApplyDuration<br/>histogram"]
  PROM -->|getPadUsersMap| PUG["padUsersGauge<br/>gauge"]
  PROM -->|expose| STATS["/stats/prometheus<br/>endpoint"]

Loading

File Changes

1. src/node/prom-instruments.ts ✨ Enhancement +36/-0

New Prometheus instruments module for hot path

• New module defining three Prometheus instruments for the scaling dive
• Exports changesetApplyDuration histogram, socketEmitsTotal counter, and padUsersGauge gauge
• Provides recordChangesetApply() and recordSocketEmit() helper functions for hot-path recording
• Separated from prometheus.ts to avoid circular dependency with PadMessageHandler

src/node/prom-instruments.ts

---

2. src/node/handler/PadMessageHandler.ts ✨ Enhancement +21/-0

Instrument message handlers with metric recording

• Import recordChangesetApply and recordSocketEmit from new prom-instruments module
• Add getPadUsersMap() function to derive per-pad user counts from sessioninfos
• Instrument handleCustomObjectMessage() to record broadcast emits
• Instrument handleCustomMessage() to record broadcast emits
• Instrument sendChatMessageToPadClients() to record CHAT_MESSAGE emits
• Instrument handleUserChanges() to measure changeset apply duration
• Instrument updatePadClients() to record NEW_CHANGES emits per socket

src/node/handler/PadMessageHandler.ts

---

3. src/node/prometheus.ts ✨ Enhancement +14/-0

Register new metrics and populate pad users gauge

• Import and register three new metrics from prom-instruments.ts with central registry
• Update monitor() function to populate padUsersGauge by calling getPadUsersMap()
• Reset padUsersGauge on each scrape to avoid stale labels for drained pads

src/node/prometheus.ts

---

View more (1)

4. src/tests/backend-new/specs/prom-instruments.test.ts 🧪 Tests +47/-0

Smoke tests for metric recording helpers

• New smoke test file verifying recordSocketEmit() increments counter by message type
• Test verifies fallback to "unknown" label when type is undefined
• Test verifies recordChangesetApply() observes duration in histogram
• Tests reset metrics before each case to ensure isolation

src/tests/backend-new/specs/prom-instruments.test.ts

---

[qodo-free-for-open-source-projects]

Code Review by Qodo

🐞 Bugs (0) 📘 Rule violations (0) 📎 Requirement gaps (0)

1. New Prometheus metrics lack flag ✓ Resolved 📘 Rule violation ⚙ Maintainability

Description

The PR introduces new Prometheus metrics (etherpad_pad_users,
etherpad_changeset_apply_duration_seconds, etherpad_socket_emits_total) that are registered and
emitted unconditionally. This violates the requirement that new features be gated behind a feature
flag and disabled by default.

Code

src/node/prometheus.ts[R26-37]

+// Added for the #7756 scaling dive: lets the load-test harness attribute
+// where time goes (apply path vs. fan-out) and confirm per-pad concurrency.
+// The metric handles live in prom-instruments.ts to avoid a circular import
+// with PadMessageHandler (which records into them on the hot path).
+import {padUsersGauge, changesetApplyDuration, socketEmitsTotal} from './prom-instruments';
+register.registerMetric(padUsersGauge);
+register.registerMetric(changesetApplyDuration);
+register.registerMetric(socketEmitsTotal);
+
client.collectDefaultMetrics({register});
const monitor = async function () {

Evidence

PR Compliance ID 9 requires new features to be behind a feature flag and disabled by default. The
diff shows the new metrics are registered in src/node/prometheus.ts and emitted/recorded from
PadMessageHandler with no flag checks or configuration gate, meaning they are enabled by default.

src/node/prometheus.ts[26-48]
src/node/handler/PadMessageHandler.ts[606-610]
Best Practice: Repository guidelines

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
New Prometheus metrics are always enabled, but compliance requires new features to be behind a feature flag and disabled by default.
## Issue Context
This PR registers three new metrics and emits them on the hot path and in the Prometheus monitor loop without any enable/disable mechanism.
## Fix Focus Areas
- src/node/prometheus.ts[26-48]
- src/node/handler/PadMessageHandler.ts[50-50]
- src/node/handler/PadMessageHandler.ts[609-609]
- src/node/handler/PadMessageHandler.ts[630-630]
- src/node/handler/PadMessageHandler.ts[671-671]
- src/node/handler/PadMessageHandler.ts[791-791]
- src/node/handler/PadMessageHandler.ts[904-904]
- src/node/handler/PadMessageHandler.ts[957-957]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---

2. Histogram includes fan-out ✓ Resolved 🐞 Bug ≡ Correctness

Description

handleUserChanges() starts etherpad_changeset_apply_duration_seconds before any processing and only
stops it in finally, after updatePadClients() completes, so the histogram includes broadcast/fan-out
time rather than isolating the apply path. This prevents the scaling-dive analysis from
distinguishing “apply is slow” vs “fan-out is slow”.

Code

src/node/handler/PadMessageHandler.ts[791]

+  const stopHistogram = recordChangesetApply();

Evidence

The timer is started at the top of handleUserChanges(), then updatePadClients() is awaited, and only
afterwards is stopHistogram() called in finally, so the histogram necessarily includes fan-out time.

src/node/handler/PadMessageHandler.ts[789-905]
src/node/prom-instruments.ts[11-15]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
`etherpad_changeset_apply_duration_seconds` is intended to time the *apply path*, but the timer currently spans the entire `handleUserChanges()` including `await exports.updatePadClients(pad)` (fan-out). This makes the metric misleading and defeats the PR’s stated purpose.
## Issue Context
- The timer is started near the beginning of `handleUserChanges()` and stopped in `finally`.
- `updatePadClients()` is awaited before the timer is stopped, so its work is included in the duration.
## Fix Focus Areas
- src/node/handler/PadMessageHandler.ts[788-905]
## Implementation notes
- Move `recordChangesetApply()` start to immediately before the “apply” work you want to measure (e.g., just before `pad.appendRevision(...)`).
- Call the returned `stopHistogram()` immediately after the apply work completes and **before** any socket emits / `updatePadClients()` fan-out.
- If you also want total end-to-end latency, keep using the existing `stats.timer('edits')` (or add a second histogram explicitly for fan-out/total).

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---

3. Unbounded metrics label ✓ Resolved 🐞 Bug ☼ Reliability

Description

handleCustomMessage() increments etherpad_socket_emits_total with msgString (an HTTP API parameter)
as the label value, allowing an API caller to generate unbounded distinct label values and
potentially exhaust memory/CPU via high-cardinality time series. Although the endpoint is API-key
protected, this still creates a sharp footgun if the key is leaked or a client misbehaves.

Code

src/node/handler/PadMessageHandler.ts[630]

+  recordSocketEmit(msg.data.type);

Evidence

The HTTP API passes an arbitrary msg string into handleCustomMessage(), which assigns it to
data.type and uses it as the type label for the Prometheus counter; prom-client counters create
distinct time series per label value.

src/node/db/API.ts[863-866]
src/node/handler/PadMessageHandler.ts[614-631]
src/node/prom-instruments.ts[17-21]
src/node/prom-instruments.ts[34-36]
src/node/handler/APIHandler.ts[180-185]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
`etherpad_socket_emits_total{type=...}` uses the message type as a Prometheus label. In `handleCustomMessage()`, that type comes directly from the HTTP API `msg` parameter, which means label cardinality is unbounded.
## Issue Context
- `sendClientsMessage(padID, msg)` passes `msg` directly into `PadMessageHandler.handleCustomMessage(padID, msg)`.
- `handleCustomMessage()` sets `data.type = msgString` and then calls `recordSocketEmit(msg.data.type)`.
- `recordSocketEmit()` uses `socketEmitsTotal.labels(type).inc()`, creating a new time series per distinct `type`.
## Fix Focus Areas
- src/node/handler/PadMessageHandler.ts[620-631]
- src/node/db/API.ts[863-866]
- src/node/prom-instruments.ts[17-21]
- src/node/prom-instruments.ts[34-36]
## Implementation notes
Pick one of:
1) **Normalize/bucket**: Map any unknown/untrusted type to a small bounded set (e.g., `API_CUSTOM_MESSAGE`, `CUSTOM`, `other`, `unknown`).
2) **Allowlist**: Only allow a fixed set of label values (e.g., `NEW_CHANGES`, `CHAT_MESSAGE`, etc.); everything else becomes `other`.
3) **Remove label at this site**: For `handleCustomMessage()`, increment with a constant label regardless of `msgString`.
Also consider enforcing a max length/character set if you keep any user-provided label value.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---