Designing Vary-aware cache keys for languages and formats

Core idea

Vary tells a cache which request fields selected a representation. If an origin varies language or encoding without naming the signal, a shared cache key can return the wrong variant.

Turn Designing Vary-aware cache keys for languages and formats into a reproducible scenario with inputs, expected state, total timeout, concurrency limit, and stop condition. The proxy is one dependency; page, browser, and test-data failures must remain distinguishable from channel failures.

What the primary source establishes

Build a language and compression cache-key contract test that detects representation mixing before production.

The primary source, RFC 9111 — HTTP Caching, RFC 9110 — HTTP Semantics, MDN Vary, MDN Accept-Encoding, MDN Accept-Language, defines the technical baseline but not every client and provider configuration. Read the normative behavior with its version and then verify your implementation. Treat anything beyond the source as a product feature that needs separate confirmation.

Controlled lab

Return ru/en and identity/gzip/br markers. Warm ru+gzip first, then request en+br and every remaining combination. Remove Accept-Language from Vary on staging to prove the negative test detects mixing, restore it, and repeat with a clean cache.

Step-by-step verification

Serve two languages and gzip or Brotli from an owned endpoint. Exercise the full header matrix through one shared cache, recording body marker, Content-Language, Content-Encoding, Vary, and Age. Reverse warm-up order as a control.

Start with one authorized URL and one proxy. Verify the exit IP, then add the target action and wait for its explicit result. Store a request ID and stage, never credentials. Add regional matrices and bounded parallelism only after single runs are stable.

Retry only proven safe reads. Respect Retry-After and back off after 429 or network bursts. Purchases, credential changes, and renewals need idempotency plus reconciliation before any repeat. A timeout does not prove failure because the external system may have completed the mutation.

• Define finite variants
• Warm one combination
• Exercise the full matrix
• Remove one Vary signal as a negative control

Evidence to retain

Keep the request-header matrix, body marker, Content-Language, Content-Encoding, Vary, Age, cache status, and warm-up order. Exclude cookies and user IDs.

Log the scenario, stage, start and finish, result code, attempt count, and correlation ID. Attach sanitized HAR or screenshots only to failures. Keep a batch summary separate from detailed rows so one failure cannot disappear among successes.

Define report columns and time format before the run. A result without context becomes a guess: the address, cache state, and changed condition are unknown. Record controlled failures as well as successes so the check proves that it can distinguish states.

Interpreting the result

Adding every field to Vary fragments the cache. Normalize a small representation set, and mark personalized responses private or no-store rather than trying to list every user signal.

Vary cannot repair incorrect origin selection and CDNs may normalize fields differently. User-Agent commonly creates excessive fragmentation.

One successful run confirms only one client, route, and moment. Repeat while changing one variable and state the limits. When observation conflicts with documentation, rule out cache, client version, and intermediaries before creating a reproducible support case.

Worked decision process

Model Designing Vary-aware cache keys for languages and formats with queued, running, succeeded, terminally failed, and uncertain states. An external timeout is uncertain because the provider may have completed the mutation. Reconcile with a read before allowing any repeat.

Limit the whole queue, each domain, each account, and retries. Add schedule jitter, honor Retry-After, and back off after bursts. A larger IP pool does not remove origin limits or infrastructure cost.

Store scenario ID, attempt, stage, timestamps, safe result code, and source task. Show stuck and uncertain work separately. Recover with one control task before releasing the bulk queue.

Common mistakes

Long sleeps hide races and immediate retries amplify incidents. Do not evade 429 by rotating addresses or run state-changing tests for one account concurrently. Wait for conditions, bound queues, isolate accounts, and use explicit terminal states.

Stop when errors rise, a source returns a limit, the task would require bypassing protection, or secrets enter logs. Save sanitized diagnostics and correct the cause first. More concurrency or another IP can hide the fault and add load without improving evidence.

Rollout and maintenance criteria

Define the decision boundary before rollout: which observation permits continuation, which requires review, and which stops the workflow. Record acceptable error ratio, maximum wait, and the owner of every exception so a temporary failure cannot silently become permanent configuration.

Review real load, cost, and quality after the first week. Schedule a small control after client, proxy-service, or network changes. Archive outdated instructions with their replacement date and reason so operators do not follow conflicting configurations.

Operational checklist

Turn the successful experiment into a short procedure covering owner, safe configuration, limits, and stop conditions. Every run needs a terminal status. After browser, library, or network changes, run a small control before the main queue.

• Success and stop are defined
• Exit IP is verified
• Waits observe events
• Retries are bounded
• Mutations are idempotent
• Artifacts contain no secrets