Org Skills

AEM performance diagnostics — tools, CDN log analysis, load testing

Supporting material for aem-performance. Agents load it on demand; it ships inside the skill folder.

RawSource

Verified: 2026-09 against https://experienceleague.adobe.com/en/docs/experience-manager-cloud-service/content/implementing/using-cloud-manager/manage-logs and https://experienceleague.adobe.com/en/docs/experience-manager-learn/cloud-service/caching/cdn-cache-hit-ratio-analysis

Open this when you need evidence: which tool produces which number, how to compute a hit ratio per path, and how to run a load test that means something.

Tool per question

Question Tool Where
Which URLs are slowest? request.log + rlog.jar Cloud Manager → Download Logs (Publish), or aio cloudmanager:environment:tail-log
Why is one request slow? Developer Console → Status (OSGi, Sling Models) and Queries Cloud Service Developer Console per environment
Is my query indexed? Explain Query / Query Performance Developer Console → Queries (Cloud); Tools → Operations → Diagnosis (6.5)
Where is the time at the edge? CDN logs (ttfb, cache), Server-Timing response header Cloud Manager → Download Logs → CDN row
Are threads blocked? 3–5 thread dumps, 5–10 s apart Adobe support or the local SDK; compare stacks across dumps
Is memory leaking? GC logs, heap dump Adobe support for production; reproduce on the SDK where possible
Field Core Web Vitals CrUX / a RUM script Not a server-side measurement — never judge LCP from Lighthouse alone
Capacity JMeter / Gatling / k6 against stage, with the CDN in the path Coordinate large tests with Adobe support per your program's policy

Application performance monitoring (for example New Relic) is entitlement-dependent. Check what your program actually has provisioned before designing a plan around it; the log surfaces above are always available.

CDN log analysis

CDN logs are JSON, one object per request, and the format is fixed. The fields that matter:

Field Use
cache HIT, MISS or PASS — PASS means the response was explicitly marked uncacheable
ttfb Time to first byte in milliseconds
url, host, method Group by these
status Separate 404 storms from real traffic
res_age How long the object had been cached
res_ctype Split HTML from images and clientlibs — they have different targets

Hit ratio for HTML by path prefix, from a downloaded log:

jq -r 'select(.res_ctype|test("text/html")) | "\(.cache) \(.url)"' cdn-*.log \
  | awk '{split($2,p,"/"); key=p[2]"/"p[3]"/"p[4]; tot[key]++; if($1=="HIT") hit[key]++}
         END{for(k in tot) printf "%-40s %5d req  %5.1f%% hit\n", k, tot[k], 100*hit[k]/tot[k]}' \
  | sort -k2 -rn

Read it as three separate problems:

  • A high PASS share means something is setting Cache-Control: private/no-store, or a cookie is on the path — an application bug.
  • A high MISS share with low res_age means over-invalidation (statfileslevel, a publish loop) or cache-key explosion (query strings, unbounded selectors).
  • A low ttfb on MISS and a high one on HIT means you are reading the wrong tier — check that you filtered to the right host.

LogForwarding config-pipeline configs push the same logs to a destination continuously, which is how you get this as a dashboard rather than an ad-hoc download.

Reading request.log

java -jar rlog.jar -n 20 request.log          # 20 slowest requests
java -jar rlog.jar -n 20 -m 1000 request.log  # slowest, ignoring anything under 1 s

Pair every slow URL with its cache value from the CDN log for the same window. A slow URL that is always a HIT at the edge is not a publish problem.

Oak query limits

Two Oak settings decide when a bad query becomes an exception instead of a warning: a read limit and an in-memory result limit (queryLimitReads, queryLimitInMemory on the QueryEngineSettings MBean). On Cloud Service these are platform-managed — you cannot raise them, which is deliberate: a query that trips them is a missing index, not a tuning opportunity. Treat Traversed N nodes in error.log as a failing test.

Load testing that means something

  1. Run against stage, which mirrors prod, with the CDN and dispatcher in the path. A test that hits publish directly measures nothing you ship.
  2. Use a realistic URL mix, including the long tail — a test that hammers one page measures your cache, not your application.
  3. Warm the cache the way real traffic does, then report cached and uncached numbers separately.
  4. Model authoring load separately on the author tier: concurrent editors, asset uploads, rollouts.
  5. Record the CDN hit ratio the test achieved next to the latency numbers. Without it the result is not reproducible.