Background
This year's annual plan includes an effort to increase logged-in readership (see DE3.1 on Meta-Wiki). For this to be feasible at a technical level, performance improvements are needed across the stack.
We started planning for this last year (see WE6.5.1 Plan work for logged-in users performance on Meta-Wiki). Between SRE and MwEng teams we arrived on the following. We drafted through various meetings meeting notes (public) and a drafting excercise (private), which I then used as my closing comment on WE6.5.1 in Asana (private):
We've converged on a two-prong approach:
- Try to increase MediaWiki backend capacity for logged-in pageviews.
- Increase effective capacity through reducing non-pageview workloads. Such as by improving CDN caching of MediaWiki API responses (thus freeing up capacity for pageviews).
- Increase effective capacity through making MediaWiki serve pageviews faster. Such as through improvements to Parsoid, ParserCache, and MediaWiki more broadly (T302623).
- Implement SLO for backend pageview timing (latency for end-users), and measure available infrastructure capacity (MW-on-k8s) for pageviews. See also Supported logged in users capacity metric (public) by Giuseppe.
- Try to serve some logged-in pageviews from the edge.
- See also Logged-in page view cacheability hurdles (restricted).
Scope
This is a tracking task for this year's work under key result ST6.4. For more details, check FY2026-2027#ST6 on Meta-Wiki:
Key Result ST6.4: By the end of Q3, our platform is able to sustain 20% of page views from logged-in users.
Including:
- ST6.4.1 Implement SLO for backend pageview request latency (user experience)
- ST6.4.2 Reduce workload from non-pageview requests
- ST6.4.3 Allow use of ESI for composition based on user ID at the edge
- ST6.4.4 Implementation of endpoints in MW to provide the ESI fragments
This task is for tracking work within the ST6.4 KR. For tracking work in other projects that impacts this work (e.g. by assuming something about cacheability, see FY26-27 ST6.4 - Logged-in performance .