Endpoint acceptance defaults

For #57, the requester chose a 10,000-row local baseline on 2026-10-20 and delegated the remaining defaults. These are starting assumptions for a small analytical service. Production availability and network targets below still require deployment validation.

Dimension Default
Result 10,000 rows; int64 id, int32 group, float64 value, 64-character UTF-8 label
Payload About 88 Arrow bytes/row, 0.84 MiB/result, excluding transport overhead
Readers 16 simultaneous full-result readers; default buffering policy
Query mix Full scans plus aggregate, join, residual sort, distinct and window checks
Placement Clients, Flight and Redis in one region/zone; target RTT at most 2 ms per hop
Initial deployment Two Flight replicas, each 4 vCPU/8 GiB RAM; separate Redis with Search; sticky routing for cookie sessions
Availability objective 99.9% successful eligible requests over 30 days; scheduled maintenance included; capacity refusals count as failures at the promised load
Latency objective At 16 readers, p95 first batch within 1 second and full result within 2 seconds; caller deadline 10 seconds
Capacity policy Buffer by default; use reject mode only with client backoff and a lower explicitly advertised admission capacity

Latency/availability are proposed operational objectives, not measured guarantees. Validate at least 100 requests per scenario for p95 and a sustained soak before using these in a service agreement. The local three-repeat check uses the maximum observed latency, not a statistically meaningful p95. Start investigating above 2 GiB Flight RSS per replica on this fixture; sampled RSS is not a memory limit. Observe Redis memory/cursors and client memory separately. A 4 vCPU deployment must be tested independently of the larger development machine.

Reproduce the local baseline

Use a disposable Redis Enterprise cluster with Search, verified TLS and username/password authentication, and an enterprise Flight server at localhost:18815 configured through its exported authenticated rediss:// REDIS_URI. See integration setup. Install the pinned dependencies from examples/python/requirements-smoke.txt. Start a fresh server per profile to avoid retained heap from prior larger workloads; supply its PID to sample RSS.

python examples/python/benchmark.py --server-pid "$FLIGHT_PID" \
  --policy-label buffer --output benchmarks/local-endpoint.json

The defaults are 10,000 rows, label width 64, 3 repeats and 16 readers. Each run creates unique fixtures and removes them. Use a separate fresh server with --stream-overflow reject, then run the same script with its URI/PID and --policy-label reject --cases sql_full_scan --repeats 1 to measure refusals. The policy label records configuration; it does not configure the endpoint.

Raw evidence, versions, source/binary fingerprints and machine details are in the local records. Flight RPC and SQL use the same flat fixture. The RESP comparison uses SCAN plus pipelined HMGET and Arrow conversion, and applies only to the unordered scan. Timings include Python checksum work and use warmed connections. All processes share one development host; no WAN delay or packet loss was emulated. RSS samples every 20 ms can miss peaks, and exclude Redis/Docker memory. Query order and retained heap affect RSS. No production sizing or general speedup follows from these results.

Failure and rollout acceptance

Run examples/python/failure_probe.py with an explicitly named disposable Redis container, a separately running rollback endpoint and an output path. It checks a one-second request deadline during a backend pause, dataset and SQL recovery after unpause, and old-reader compatibility with newly written data. gRPC may surface the expired deadline as timeout or cancellation; the probe records the observed type and bounds its elapsed time. This is a backend outage probe, not Redis failover or an availability measurement.

The Go race suites cover slow consumers, cursor/pool cleanup, cancellation across replicas, lost sessions and restart. Lost cookie sessions must fail before SQL execution; clients reconnect and restore settings/temp objects. Upgrade ticket readers together: old readers ignore new cancellation records. On rollback, discard query tickets and cookie sessions and verify SQL plus dataset reads. Storage format compatibility does not imply active session compatibility.

Before production acceptance, exercise the assumed 2 ms network and proposed 4 vCPU/8 GiB instances, sustained load, replica draining/restart, actual Redis failover and the deployment system's rollback. Monitor RPC errors/latency, health, pool/stream saturation, result refusals, stalled results, Redis cursors and process memory. #57 remains open for this deployment evidence; the workload and starting objectives no longer require further user input.

Recorded local result (2026-10-20)

Measured binary revision 1fa8716, Redis 8.6.2, on an 18-logical-CPU macOS host with 36 GiB RAM. The raw records pin the Redis ADBC revision, client packages, source manifest and binary hash. Each profile used a fresh Flight process.

Path Median first batch Median full result
Flight RPC scan 6.63 ms 6.82 ms
Flight SQL scan 55.07 ms 55.79 ms
RESP SCAN/HMGET plus Arrow conversion 7.57 ms 73.03 ms

These are three repeats per path. All recorded single-query cases and successful concurrent queries met the local 1-second first-batch/2-second total checks. The full-mix profile's 16 SQL readers completed in 0.498 seconds, with sampled server/client peaks of 397.4/197.6 MiB during concurrency. The separate scan-only buffer profile completed 16/16 readers in 0.429 seconds (211.4/200.7 MiB peaks). The reject profile accepted 11 and refused 5 with RESOURCE_EXHAUSTED; the exact split is scheduling dependent and is not a 16-reader service guarantee. A paused Redis request timed out in 1.004 seconds, then SQL and dataset reads recovered. The pre-change 4120689 server read newly written data successfully.