course-update-recon
Skill431 starsChanged 8 months agoMITView on GitHub
Characterize a slow / stuck course-update worker on dashboard.wikiedu.org or outreachdashboard.wmflabs.org. Use this skill when asked to investigate why a course update is taking unusually long, hung, or running for many hours, or to estimate the scale-of-work of an in-flight update. Pure HTTP recon against public APIs — never SSH into prod.
In this repository7 other files it gives its agents.
- AGENTS.mdAGENTS.md
- .claude/CLAUDE.mdCLAUDE.md
- commitSkill
- feature-specSkill
- google-doc-change-requestsSkill
- prepare-prSkill
- review-prSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 47
last completed: update — if a worker is currently running, it won't appear here, but flags.unfinished_update_logs shows the start_time of …
- Very specific: line 98
Use benchmarks/probe_course_contributions.rb — it runs the recon against list=usercontribs for ns=0 on each non-wikidata wiki, …
- Very specific: line 122
Use benchmarks/probe_timeslice_gap.rb — it queries the same endpoint the updater uses (revisions.php on replica-revision-tools.wmcloud.org) …
- Distinctive: line 166
If projected « observed: it's a service-availability problem, not a scale problem.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Course-update recon
Workflow for characterizing a slow or stuck course update without ever touching prod. The goal is a per-axis estimate of the work-in-flight (or work-just-completed) so we can predict runtime and pick a fix.
Hard rule: never SSH to peony-web, peony-database, peony-sidekiq, dashboard.wikiedu.org, or any other prod host.* All recon is via public dashboard JSON endpoints + public Wikimedia APIs. If you need something only available inside the prod app, ask the user to run a one-liner and paste the result back.
Step 1 — Resolve course id → slug
The numeric id is what the user usually has; the public JSON endpoints
key on slug. The /find_course/<id> endpoint redirects from id to the
canonical course URL — read the Location: header to get the slug:
curl -sSI -A "WikiEdu-Dashboard-Recon/1.0 (sage@wikiedu.org)" \
https://outreachdashboard.wmflabs.org/find_course/<ID> | grep -i location
# Location: https://outreachdashboard.wmflabs.org/courses/<SLUG>For Wiki Ed Dashboard courses use https://dashboard.wikiedu.org/find_course/<ID>.
Slugs may contain ', : etc. — URL-encode before hitting the JSON
endpoints.
Step 2 — Pull course.json + users.json
SLUG_ENC="...url-encoded slug..."
HOST=https://outreachdashboard.wmflabs.org # or dashboard.wikiedu.org
curl -sS "$HOST/courses/$SLUG_ENC/course.json"
curl -sS "$HOST/courses/$SLUG_ENC/users.json"From course.json, note: start, end, home_wiki, wikis,
updates.average_delay, and updates.last_update (start_time,
end_time, processed, reprocessed). The last_update.end_time is the
last completed update — if a worker is currently running, it
won't appear here, but flags.unfinished_update_logs shows the
start_time of any in-flight (or orphaned) update, so "how long has the
current run been going" is answerable. flags.update_logs holds the
last ~10 completed updates — compute their durations; a uniform
duration across runs means steady-state cost, not a one-off stall.
processed / reprocessed count TIMESLICES, not revisions (one
per handle_timeslice leaf in UpdateCourseWikiTimeslices). Each
processed timeslice costs one revisions.php query to
replica-revision-tools per wiki, whether or not it contains any
revisions.
From users.json, extract users[].username (filter by role —
role 0 = student, 1 = instructor). Only role-0 students are tracked
for revisions (CourseRevisionUpdater uses course.students), so
recent edits by instructors/facilitators are irrelevant to update cost.
Step 3 — Identify the cost shape
Two distinct shapes produce long updates; pick the probe accordingly:
- Revision-volume shape (active course, many edits): cost is
dominated by per-revision work (ref-counter, Lift Wing, diff
analysis). Tell:
processedis modest but durations scale with recent edit volume. → Step 4. - Empty-timeslice-scan shape (dormant course, end date far in the
future): cost is dominated by scanning empty daily timeslices, one
replica query each. Tells:
processed≫ the course's lifetimeedit_count; every update inupdate_logshas near-identical duration andprocessedcount;processedgrows by exactly (number of tracked wikis) per day. → Step 5.
Mechanism of the second shape: TimesliceManager#get_ingestion_start_time_for_wiki
pins the ingestion start at the timeslice containing the latest
tracked student revision (max last_mw_rev_datetime). Empty
timeslices never advance that watermark (process_timeslices is
skipped when there's no new data), so every update re-scans every
daily timeslice from the last student edit to today, per wiki,
forever — the window grows by 1 slice/wiki/day until course end.
Step 4 — Revision-volume shape: bucket contributions along the three cost axes
Per the 2026-04-24 benchmark (.claude/plans/benchmark_notes-2026-04-24.md):
| Axis | Per-unit cost | Driver |
| Non-wikidata (enwiki etc.) ns=0 revs | ~4 s/rev | ReferenceCounter (2× per rev) + Lift Wing |
| Wikidata revs (any ns) | ~0.2 s/rev | Replica + WikidataDiffAnalyzer (batched) |
| Commons uploads | 0.5 s (healthy) – 10 s (observed stalled) per upload | UploadImporter serial thumburl + usage_count |
Use benchmarks/probe_course_contributions.rb — it runs the recon
against list=usercontribs for ns=0 on each non-wikidata wiki,
list=usercontribs (all ns) on wikidata, and list=allimages on
Commons, then projects total runtime from the per-axis costs.
SLUG='Polskojęzyczna_Wikipedia/Wikiprojekt_Nauki_medyczne' \
ruby benchmarks/probe_course_contributions.rbRequired env: SLUG (course slug). Optional: HOST (default
outreachdashboard.wmflabs.org; switch to dashboard.wikiedu.org for
Wiki Ed courses), WINDOW_START / WINDOW_END to narrow the probe
window, PER_USER_CAP to raise the per-user pagination cap (default
5000), USERS_LIMIT to spot-check a subset.
Gotcha: list=usercontribs with multiple ucuser values returns
per-user blocks, NOT timestamp-merged results, and caps at 50 users —
never use a multi-user usercontribs query to find "the latest edit by
any of these users". Query the replica endpoint instead (below), which
also matches the updater's own tracked-namespace filtering.
Step 5 — Empty-timeslice-scan shape: probe the watermark gap
Use benchmarks/probe_timeslice_gap.rb — it queries the same
endpoint the updater uses (revisions.php on
replica-revision-tools.wmcloud.org) to find each wiki's watermark
(latest tracked student revision), computes the expected
timeslice-scan count, times empty-day queries for the per-slice unit
cost, and projects runtime:
SLUG='tlv_university/TLV_University,_Science_Oriented_Youth_-_Alpha_Program_-_Cycle_C_(2016-2017)' \
ruby benchmarks/probe_timeslice_gap.rbExpected scans = Σ over wikis of (days from the timeslice containing
the last student edit, through today, inclusive; slice boundaries are
aligned to the course start time-of-day). If this matches
last_update.processed exactly (small positive excess = split
timeslices), the diagnosis is confirmed. Calibration (course 10781,
2026-08-20): ~0.25–0.30 s per empty-slice replica query + ~35%
Rails-side overhead ≈ 0.35 s/slice end-to-end; 3,867 slices → 22.9 min
per update, ~10×/day.
Querying the replica endpoint directly (what the script does):
https://replica-revision-tools.wmcloud.org/revisions.php?lang=he&project=wikipedia&usernames[]=A&usernames[]=B&start=YYYYMMDDHHMMSS&end=YYYYMMDDHHMMSSResponse is {"success":true,"data":[...]} with rev_timestamp
fields. For wikis with nonstandard db names use db= instead of
lang=/project= (wikidata → db=wikidatawiki, commons →
db=commonswiki; full list in Replica::SPECIAL_DB_NAMES). A
malformed/unknown project silently falls back to a bogus db and
returns a 502 "Cannot connect to database" error — check
success:false bodies rather than assuming the service is down.
Step 6 — Sanity-check projection vs. observed runtime
Compare projected wall time to:
- The user's reported runtime ("update has been running 9 hours")
updates.average_delay(seconds between completed updates) — if the projection is way under average_delay, an external factor (Toolforge service down, Lift Wing slow, etc.) is likely the culprit, not the course's edit volume.
If projected « observed: it's a service-availability problem, not a scale problem. If projected ≈ observed: it's pure scale and a reference-counter / Lift Wing batch fix would help.
Step 7 — Report
Summarize for the user:
- Course shape (home_wiki, # users, window, last completed update)
- Which cost shape it is, and the per-axis / per-wiki numbers
- Projected runtime range (low/high) vs. observed
- Whether projected runtime explains observed runtime, or whether a different bottleneck is implicated (and which one to investigate next — Toolforge tool status, Lift Wing latency, etc.)
- For the empty-timeslice-scan shape, also report the daily waste (updates/day × duration) and growth rate — the interesting question becomes "how many other courses have this shape", since every dormant course with a far-future end date burns worker time on every cycle until its end date.
Reference files
benchmarks/probe_course_contributions.rb— revision-volume probe; drives offSLUGenv varbenchmarks/probe_timeslice_gap.rb— empty-timeslice-scan probe; drives offSLUGenv var; validated 2026-08-20 against course 10781 (predicted processed=3867 exactly; projected 22.1 min vs 22.9 observed)benchmarks/cuwprofile_http.rb— per-HTTP-call profiler for a real UpdateCourseStats run (dev DB only, not prod recon); supportsMODE=cold|warm|incremental|cold+warm.claude/plans/benchmark_notes-2026-04-24.md— cost model derivation + per-endpoint timings.claude/plans/plan_course_update_perf_regression-2026-04-24.md— loose ends from the 35819 investigation (UploadImporter, sidekiq-status TTL, etc.)
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

