Antalya 26.6: Iceberg: support external paths in tables

[zvonand]

Dropped from this backport: the AI dropped these surfaces rather than pulling in a missing prerequisite. Reviewers: confirm each is genuinely optional.

• manifest-only compaction integration (compactIcebergManifests, writeConsolidatedManifestFile, isCurrentManifestListAboveThreshold, IcebergMetadata::optimizeManifestFiles) — that feature is not on antalya-26.6, only the PR's edits to it were dropped
• query-condition-cache keying by the Iceberg metadata path (makeQueryConditionCacheKey, ObjectInfo::getIdentifier(bool)/getIdentifierForPath(path, bool) overloads) — depends on upstream query-condition-cache/ETag work not on antalya-26.6
• gtest_storage_object_storage_archive.cpp update — that test file does not exist on antalya-26.6

Changelog category (leave one):

• Improvement

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

Support Iceberg tables that have files outside table location or on different storage (ClickHouse#90740 by @zvonand).

Cherry-picked from ClickHouse#90740.

---

Closes ClickHouse#84609

Current logic is that all table files (data files, manifests, manifest lists) shall be "inside" the table location, in the same storage. This PR abandons that logic: now files can be located anywhere, even on a different storage type (e.g. all metadata is on s3, data file in in local storage).

In some cases, that old logic shoots back: even incorrect (non-existing) paths, i.e. with redundant items in a prefix were parsed in such a way that files from another locations were read (/path/to/item was read when the specified location was /path/to/to/item or vice versa)

Documentation entry for user-facing changes

• Documentation is written (mandatory for new features)

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)

[github-actions]

Workflow [PR], commit [a623cd1]

[zvonand]

@blau-ai

[blau-ai]

CI triage — 3 distinct red checks: 2 PR-caused, 1 flaky/unrelated

I looked at every failing check on a623cd1. They collapse into three real signals (the top-level PR check is just the rollup gate, and each failure reproduces identically on both aarch64 and x86_64, so nothing here is a runner one-off):

Check	Verdict	Root cause
Regression Iceberg (2) / iceberg_2	🔴 PR-caused	Iceberg export/write path
Regression S3Export (part) / s3_export_part	🔴 PR-caused	same write path
Stateless tests (arm_binary, parallel)	🟡 flaky / unrelated	03518_alter_logical_race

The good news first: the reader side this PR actually targets is healthy. On the base branch iceberg_2 fails 8 iceberg engine (read) scenarios; on this PR those are green (0 read failures). The intended improvement (external/relative Iceberg paths on read) appears to work. The regressions are all on the write side.

---

🔴 1. Iceberg (2) / iceberg_2 — PR-caused

• PR: 242 leaf scenarios failed, and every one is under /iceberg/export partition (842 [Fail] lines incl. parent suites; the rest of the module passes). Fails on both arches identically.
• Base branch (antalya-26.6, latest MasterCI 31713806632): only 14 iceberg_2 failures total, of which just 4 touch export partition.
• Even the smoke test .../sanity/export single partition fails, across no catalog / ice catalog / glue catalog and both plain/replicated merge tree → the EXPORT PARTITION operation itself is broken, not a single edge case.

Classification: whole-suite regression, base→PR = 4→842. This is the PR.

🔴 2. S3Export (part) / s3_export_part — PR-caused (same root cause)

• PR: 66 (release) / 68 (arm) [Fail], spread across export part/{concurrent alter, concurrent other, error handling, pending mutations and patch parts, rbac, table functions}.
• Base branch: only 5 failures, all in one sub-suite (feature support validation).
• Same story: base→PR = 5→66. Both features share the Iceberg manifest/data-file writing code this PR reworks.

🟡 3. Stateless tests (arm_binary, parallel) — flaky, not this PR

• Exactly one test failed: 03518_alter_logical_race [ FAIL ] 182.81 sec — a concurrency/timing stress test (ALTER race) with a documented history of stabilization commits (Let 03518_alter_logical_race extend its budget…, Stabilize 03518_alter_logical_race…).
• This PR touches nothing in that path (Iceberg/object-storage only). Safe to re-run.

---

Where the write regression most likely lives

This backport of upstream ClickHouse#90740 re-plumbed the Iceberg write path to support files on secondary/external storages. Every manifest/data-file path is now routed through the new resolveObjectStorageForPath(...) (src/Storages/ObjectStorage/Utils.cpp:962), and a SecondaryStorages is threaded through generateManifestList / writeMetadataFiles / IcebergStorageSink / Iceberg::mutate:

• IcebergWrites.cpp:929-935 — previous-snapshot manifest list now resolved via resolveObjectStorageForPath instead of the old direct path_resolver.resolve(manifest_list).
• Mutations.cpp — writeMetadataFiles / mutate now require SecondaryStorages &.

For the ordinary single-storage export (no external files — which is exactly what these tests do), that resolver must fall back to the base storage. If it instead throws (e.g. resolveObjectStorageForPath → tryResolveObjectStorageForPath reaches Utils.cpp:845 "Unsupported storage scheme" for a serialized relative path that doesn't start with /), or if the manifests are now written with absolute paths the test's pyiceberg readback can't resolve, that breaks every export — matching the wholesale failure.

⚠️ I couldn't fetch the exact DB::Exception text: the regression console only carries the [Fail]/AssertionError summary; the real message lives in the S3 report.html, and web fetch is disabled in my environment. The fastest next step is to open the report and grab the Code: NNN … DB::Exception from the first .../sanity/export single partition failure — that will point straight at the throwing line:

• iceberg_2: https://altinity-build-artifacts.s3.amazonaws.com/REFs/2154/merge/a623cd16259d0d213bf8d0e599804c9318564f8c/regression/x86_64/with_analyzer/zookeeper/without_thread_fuzzer/iceberg2/report.html
• s3_export_part: https://altinity-build-artifacts.s3.amazonaws.com/REFs/2154/merge/a623cd16259d0d213bf8d0e599804c9318564f8c/regression/x86_64/with_analyzer/zookeeper/without_thread_fuzzer/s3part/minio/report.html

Suggested fixes

1. Confirm the exact exception from the report (above) — that decides between the two hypotheses.
2. Make the write path's resolveObjectStorageForPath fall back cleanly to the base storage for serialized relative metadata paths (the single-storage case), rather than throwing / emitting absolute paths. Compare the serialized form fed at IcebergWrites.cpp:932 and Mutations.cpp against what the resolver's relative-path branch (Utils.cpp:577) expects.
3. Re-check whether any write-path caller passes an empty/half-populated SecondaryStorages where the read path populates it.

I did not push anything. Want me to open a blau/* PR with a proposed fix once we have the exact exception, or do you want to drive it? (I can't build/run ClickHouse here, so any fix I write would be validated by re-running these two regression suites on the PR.)

_{Evidence: PR run 31795945657 (jobs iceberg_2 94806921483/94805671736, s3_export_part 94806921454/94805672014, stateless 94777903141); base run 31713806632 (jobs 94649901976, 94649902015). I can't build/test locally — this is static + CI-log analysis.}