Antalya 26.8: auto-grp-pr-2378: combined port of 2 PRs

[zvonand]

Changelog category (leave one):

• Improvement

Changelog entry (a user-readable short description of the changes that goes to CHANGELOG.md):

Add an Antalya-specific protocol version that the server advertises in the ServerHello and the client advertises in the Hello. It uses no upstream protocol-version slot and still works with older clients and servers.
Add an Antalya-specific protocol version that the server advertises in the ServerHello and the client advertises in the Hello. It uses no upstream protocol-version slot and still works with older clients and servers (#2378 by @zvonand, #2451 by @zvonand).

CI/CD Options

Exclude tests:

• Fast test
• Integration Tests
• Stateless tests
• Stateful tests
• Performance tests
• All with ASAN
• All with TSAN
• All with MSAN
• All with UBSAN
• All with Coverage
• All with Aarch64
• All Regression
• Disable CI Cache

Regression jobs to run:

• Fast suites (mostly <1h)
• Aggregate Functions (2h)
• Alter (1.5h)
• Benchmark (30m)
• ClickHouse Keeper (1h)
• Iceberg (2h)
• LDAP (1h)
• Parquet (1.5h)
• RBAC (1.5h)
• SSL Server (1h)
• S3 (2h)
• S3 Export (2h)
• Swarms (30m)
• Tiered Storage (2h)

Combined port of 2 PR(s) (group auto-grp-pr-2378). Cherry-picked from #2378, #2451.

• Antalya 26.6: Add an Antalya-only protocol version advertised in the ServerHello #2378 — Antalya 26.6: Add an Antalya-only protocol version advertised in the ServerHello
• Advertise the Antalya protocol version in the client Hello #2451 — Advertise the Antalya protocol version in the client Hello

[github-actions]

Workflow [PR], commit [179c3f7]

[zvonand]

@blau-ai

[blau-ai]

CI triage

Verdict: CI is red, but 0 of the failures are caused by this PR. Every failing check is pre-existing / infra / flaky, and I can show the same failures on an unrelated PR against the same base. The PR's own relevant surface (builds, fast test, stateless amd/arm, integration incl. distributed-plan, stress, fuzzers) is green.

Key cross-check: #2488 (a transaction-metadata change that touches none of the protocol code) shows the identical Grype results, the identical "Parser memory check 1/2", and the identical swarm failure counts/scenarios — so these live on antalya-26.8, not here.

Check	Verdict	Why
Grype Scan (keeper / server / server-alpine)	not PR-caused — infra	Base-image CVE scan
Parser memory check	not PR-caused — infra	Missing master baseline artifact
Regression swarms (aarch64 + release)	not PR-caused — pre-existing	Connection/socket-timeout swarm tests, same on #2488
Stateless tests (arm_binary, parallel)	not PR-caused — flaky	1 unrelated flaky test, 21/21 reruns passed, non-blocking

---

Per-failure detail

1. Grype Scan — keeper (4 high/critical), server-alpine (1 high/critical), server (0)
These are vulnerability scans of the published Docker images, i.e. CVEs in the base OS/packages. This PR only changes C++ handshake code and some tests — it cannot add or remove image CVEs. The identical counts appear on #2488. Pre-existing/infra; not PR-caused. Next step: handle via the usual image-CVE process, independent of this PR.

2. Parser memory check — ERROR
Sub-step Resolve master binary → No master 'clickhouse-examples' artifact was found. The job needs the master/baseline clickhouse-examples build to diff memory against, and that artifact wasn't available. Same ERROR on #2488. Infra/dependency, not a regression. Next step: safe to re-run once the baseline artifact exists; nothing to fix in the PR.

3. Regression swarms (aarch64 + release) — 3 failed scenarios
The failed scenarios are:

• cluster discovery/check cluster discovery with wrong cluster name — ALL_CONNECTION_TRIES_FAILED (expected exit 23, got 33)
• node failure/cpu overload — SOCKET_TIMEOUT (receive timeout 300000 ms) While executing Remote
• node failure/cpu overload all swarm nodes — SOCKET_TIMEOUT … While executing Remote

These touch "connection"/"remote" in the error text, which is worth a careful look given this PR edits the handshake — but the evidence says it's not this PR:

• The exact same 3 scenarios fail identically on Antalya 26.8: Retry transaction metadata writes under noexcept callbacks instead of terminating #2488, which doesn't touch the protocol → they predate this PR on antalya-26.8.
• The change here is purely additive to the Hello name string (append/parse " (antalya:N)"); it doesn't alter replica selection, timeouts, or error codes, and an unmarked peer is handled gracefully (parseMarker → 0).
• Integration tests (arm_binary, distributed plan) passed 0/7614 — those exercise exactly the server↔server remote handshake this PR modifies. If the marker broke remote connections, they'd fail in bulk.
• The PR's new test_antalya_protocol (negotiation over remote(), including against an old unmarked node) is in that green integration run.

The two cpu overload scenarios are socket timeouts under deliberate CPU stress — classic infra flakiness. Pre-existing/infra; not PR-caused. Next step: safe to re-run; track with the swarm-regression flakiness, not this PR.

4. Stateless tests (arm_binary, parallel) — Failed: 1
04218_limit_by_partitions_independently — an EXPLAIN PIPELINE diff (Resize/ReadPool,Thread vs Concat/ReadPoolInOrder,InOrder), i.e. MergeTree read-order nondeterminism. CI labeled it flaky and its own rerun diagnosis passed 21/21 ("not reproducible, likely transient"). The check is marked non-blocking (do_not_block_pipeline_on_failure). The test is about LIMIT BY partition reading and has nothing to do with the protocol handshake. Flaky; not PR-caused. Next step: re-run / ignore.

---

PR health check

The change is clean and well-scoped: it adds a self-contained DBMS_ANTALYA_PROTOCOL_VERSION advertised as a " (antalya:N)" suffix in the Hello/ServerHello name, parsed back with min(own, peer) and ignored by non-Antalya peers (parseMarker → 0). The compatibility-sensitive paths all passed: Fast test, every Build variant, stateless amd_debug (parallel + sequential) and arm_binary sequential, integration tests incl. distributed-plan (remote handshake) with 0 failures, stress tests, and both AST/BuzzHouse fuzzers. The new unit test (gtest_antalya_protocol) and integration test (test_antalya_protocol, which covers new↔old node negotiation and the unmarked-worker case) back up the backward-compat claims. I'd consider this green on the merits — the red checks are all base-branch/infra noise. Re-running the swarms and Parser-memory jobs is the only cleanup, and neither needs a code change.

_{🤖 Automated triage by @blau-ai. Evidence: praktika result_pr.json, swarms report.html, and a cross-check against #2488. I can't build/run ClickHouse — correctness is ultimately confirmed by CI.}