A response cache should be cheap to refill. Our cache entries shared Valkey’s append-only log with catalog metadata and writer coordination, so filling a cold response also added durable log traffic. We changed only response publication to skip incremental AOF propagation. Catalog updates, generations and writer tokens retain their required durability.

In a paired experiment, enabled cold-query p50 improved about 21% at both one and eight clients. Eight-client p95 fell 67.8%, from 43.949 to 14.169 ms. Native engines still handled these cold queries faster, and warm results included regressions. The improvement is specific to this cache-fill workload.

Disposable entries, durable authority

The publication script validates the deployment, generation and quiet writer state before storing a bounded response with an expiry. It disables propagation only around that response write, then restores normal propagation. Administrative claims, catalog publication and writer admission/release keep the existing appendfsync always requirements.

This does not mean response entries can never survive a restart. An AOF rewrite or snapshot can capture entries already in memory. A restored entry must still pass identity, build/configuration, generation, expiry and writer-state checks. An old response cannot become current merely because a rewrite saved it. The deployment guide explains these boundaries. This is not a guarantee that every fill avoids all fsync activity.

Cold-query results and log growth

We compared baseline 04b6d7e with measured candidate a778a35, using old, new, new, old order in separate serving-enabled and serving-disabled profiles. These measurements describe those builds, not a later merge. Values are median trial quantiles in milliseconds.

Gateway cold sorted filter p50 old → new p95 old → new
Serving enabled, one client 1.943 → 1.537 2.277 → 1.715
Serving enabled, eight clients 4.578 → 3.614 43.949 → 14.169
Serving disabled, one client 1.353 → 1.359 1.515 → 1.562
Serving disabled, eight clients 3.336 → 3.150 44.380 → 50.073

The enabled cold profiles added 9,103,721 and 9,103,723 bytes to the baseline AOF, compared with 2,016 bytes in each candidate round. Disabled cold profiles also added 2,016 bytes; warm profiles added none. No rewrite occurred within these sampled intervals, so log compaction does not account for that difference. The observations support reduced incremental log work, not a count of individual fsync calls or proof that all latency variation has one cause.

Native cold p50 controls from the enabled rounds remained lower:

Cold sorted filter Elasticsearch old → new OpenSearch old → new
One client 0.579 → 0.581 0.487 → 0.498
Eight clients 1.201 → 1.295 1.247 → 1.182

The disabled eight-client p95 worsened 12.8%, even though response fills were bypassed. Native controls also moved. Those results remain in the complete 42-row summary.

Warm results were mixed

Warm hits do not publish another response. Cells show one / eight clients, in p50 milliseconds.

Warm gateway operation Old New
Read one document 0.302 / 0.585 0.305 / 0.567
Filtered, sorted page 0.396 / 0.911 0.387 / 0.914
Page after a sort cursor 0.374 / 0.923 0.395 / 0.905
Keyword terms buckets 0.328 / 0.776 0.331 / 0.778
Two filter counts 0.321 / 0.768 0.322 / 0.709

Single-client sorted pagination worsened 5.8% at p50 and 21.7% at p95 (0.439 → 0.535 ms). Eight-client terms p95 rose 5.15% (1.210 → 1.272 ms). We do not claim a general warm-read improvement from this change.

Check crashes as well as speed

An isolated Valkey 8.1.10 proof exercised the production Lua scripts. Publishing a response produced a hit while AOF size stayed 257 → 257 bytes. The final six-check run also verified persistent unresolved claims and writers, durable catalog commit and releases, rewrite-restored entries with their original expiry, rejection after a durable generation change, and writer durability after a publication error.

The final proof and proof script are retained. So are both earlier attempts: the first failed before checks because the harness mishandled INFO JSON readiness parsing; the second passed an earlier five-check suite. They are not silently discarded or presented as equivalent to the final proof.

Replay the comparison

All 93,600 measured HTTP responses passed correctness checks: 72,000 warm, 14,400 enabled cold and 7,200 disabled cold. Each comparison uses six trial quantiles per image. The 10,000-document corpus, requests, exact totals and validation profile v3 stayed fixed. Cold means new cache keys on loaded data.

Elasticsearch 8.15.0 and OpenSearch 3.9.0 each had two CPUs and 4 GiB. Gateway, Qdrant 1.19.2 and durable Valkey shared the same aggregate budget. Qdrant retained automatic search pools. The driver consistently recreates the gateway and primes all engines outside timing. No builds or tests ran during measurement.

Follow the pinned preparation instructions. Build the baseline as qdrant-es-gateway:typed-readview and candidate as qdrant-es-gateway:ephemeral-publish, using the revisions above. Record local image IDs; rebuilt images may differ from the historical manifest. After preparation:

old_id=$(docker image inspect qdrant-es-gateway:typed-readview | python3 -c 'import json,sys; print(json.load(sys.stdin)[0]["Id"])')
new_id=$(docker image inspect qdrant-es-gateway:ephemeral-publish | python3 -c 'import json,sys; print(json.load(sys.stdin)[0]["Id"])')
python3 benchmark/shared-cache/ephemeral_round.py on-main old 1 \
  --old-image-id "$old_id" \
  --new-image qdrant-es-gateway:ephemeral-publish --new-image-id "$new_id" \
  --new-source a778a357f677345939e61b559bf6beac234624d7
# Repeat on-main new 1, new 2, old 2, retaining all flags.
# Then off-control old 1, new 1, new 2, old 2, retaining all flags.
python3 tests/cache_ephemeral_aof.py --output /tmp/cache-aof-proof.json

The manifest retains twelve raw profiles, eight resource records, all proof attempts, exact scripts and hashes. Public records omit only ephemeral container identifiers and failed-proof resource names. Resource counters span setup and native phases as well as gateway requests; they cannot attribute individual latencies. No delayed fsync increments were recorded in sampled profiles. Earlier experiments are kept separate.

For a workload that repeatedly fills the cache, disposable publication removes unnecessary durable response storage while preserving durable coordination. That benefit coexists with the native engines’ cold-query advantage and the adverse warm results above.