Prefer Use-And-Recover Over Preflight Checks In Connection Pools

Summary#

For pooled connections, extra liveness probes can add latency and still be wrong under some transports. A better default is to reuse the connection directly, detect real failure on first use, and reconnect only when the operation proves the connection is dead.

Problem#

The system paid an extra ping/pong round trip before every use of a pooled connection, which doubled small-query latency and relied on stale-connection heuristics that were not reliable across protocols like HTTPS.

Solution#

Remove unconditional pre-use pings, attempt the real request first, and reconnect on actual failure; when stale detection is unavoidable, use protocol-aware primitives such as socket or TLS peeking rather than generic readiness polling.

Failure Modes#

Keeping expensive preflight probes that dominate latency for small operations
Using transport-agnostic liveness checks that misread TLS state
Failing to make retries idempotent after reconnect
Turning intermittent disconnects into noisy user-visible failures because reconnect happens too late or not at all

Sources#

https://github.com/huggingface/transformers/pull/47181
https://github.com/huggingface/transformers/pull/46842
https://github.com/huggingface/transformers/pull/47381
https://github.com/huggingface/transformers/pull/47382
https://github.com/huggingface/transformers/pull/46854
https://github.com/huggingface/transformers/pull/47252
https://github.com/huggingface/transformers/pull/47369
https://github.com/huggingface/transformers/pull/47218
https://github.com/huggingface/transformers/pull/47032
https://github.com/huggingface/transformers/pull/47150
https://github.com/microsoft/ML-For-Beginners/pull/1002
https://github.com/microsoft/ML-For-Beginners/pull/1001
https://github.com/microsoft/ML-For-Beginners/pull/1000
https://github.com/microsoft/ML-For-Beginners/pull/994
https://github.com/microsoft/ML-For-Beginners/pull/991
https://github.com/microsoft/ML-For-Beginners/pull/990
https://github.com/microsoft/ML-For-Beginners/pull/989
https://github.com/ClickHouse/ClickHouse/pull/111002
https://github.com/ClickHouse/ClickHouse/pull/110402
https://github.com/ClickHouse/ClickHouse/pull/110068
https://github.com/ClickHouse/ClickHouse/pull/110991
https://github.com/ClickHouse/ClickHouse/pull/109746
https://github.com/ClickHouse/ClickHouse/pull/109419
https://github.com/ClickHouse/ClickHouse/pull/106894
https://github.com/ClickHouse/ClickHouse/pull/110608
https://github.com/ClickHouse/ClickHouse/pull/110590
https://github.com/ClickHouse/ClickHouse/pull/109439
https://github.com/ClickHouse/ClickHouse/pull/110910
https://github.com/ClickHouse/ClickHouse/pull/110121
https://github.com/vercel/turborepo/pull/13400
https://github.com/vercel/turborepo/pull/13398
https://github.com/vercel/turborepo/pull/13395
https://github.com/vercel/turborepo/pull/13394
https://github.com/vercel/turborepo/pull/13393
https://github.com/vercel/turborepo/pull/13390
https://github.com/vercel/turborepo/pull/13387
https://github.com/vercel/turborepo/pull/13385
https://github.com/vercel/turborepo/pull/13386
https://github.com/vercel/turborepo/pull/13392
https://github.com/vercel/turborepo/pull/13391
mined_at: 2026-07-20T00:10:25Z

Sagwan Revalidation 2026-07-20T00:49:33Z#

verdict: ok
note: 핵심 권장안은 현재도 통용되며 지연·재시도 caveat도 적절함