Fix option-4 empty-prefix LIST regression; add window-function LIST_ALL

[zichengl]

Summary

Two related changes for the option-4 (window-function) LIST path introduced in #3260:

1. NamedBlobPath.parseS3: collapse empty-string prefix to null. S3 clients can send prefix= with no value, which is semantically equivalent to omitting the prefix entirely. Routing the empty case to LIST_ALL_QUERY avoids triggering LIST_WITH_PREFIX_SQL with blob_name LIKE '%', which under option 4 produces a Window-over-full-container-scan plan that can exceed MAX_EXECUTION_TIME on large containers.
2. MySqlNamedBlobDb: gate LIST_ALL_QUERY behind the same listNamedBlobsSQLOption knob as LIST_WITH_PREFIX_SQL. Options 2/3 keep the existing INNER-JOIN + MAX-grouped subquery shape; option 4 adds a window-function variant matching the option-4 with-prefix plan shape and semantic.

Default listNamedBlobsSQLOption remains 2. Option 2 and 3 deployments are byte-identical to master.

Motivation

After deploying option 4 (#3260 shipped in 0.5.163) downstream, a load/correctness suite failed S3ListTests.testSuccessListObjectsEmptyPrefix against a container with a substantial number of named blobs. All other S3 LIST tests passed including the prefix-filtering tests that directly exercise option 4's new SQL — the only failure was the empty-prefix case.

Tracing through the OSS code path: S3 ListObjects with explicit prefix="" arrives at NamedBlobPath.parseS3 as blobNamePrefix = "" (not null), then through NamedBlobListHandler.listRecursivelyInternal into MySqlNamedBlobDb.run_list_v2:

String queryStatement = blobNamePrefix == null ? LIST_ALL_QUERY : LIST_WITH_PREFIX_SQL;

"" is not null, so the empty case routes to LIST_WITH_PREFIX_SQL. With prefix="", the V4 binder produces parameters ("%", ""), which under option 4 plans as:

• Single PK range scan over the entire (account_id, container_id) prefix
• Window(MAX(version) OVER (PARTITION BY blob_name)) over the full result
• Outer Selection + ORDER BY blob_name LIMIT N

On a large enough container this exceeds a typical 5s MAX_EXECUTION_TIME server kill, returning 500 to the S3 SDK after 3 retries. Pre-#3260 with option 3, the same empty-prefix request also routed to LIST_WITH_PREFIX_SQL but used the correlated-subquery shape, whose per-row probe pattern apparently completed within the kill threshold on the same dataset.

The OSS integration matrix in #3260 (MySqlNamedBlobDbListOperationIntegrationTest) only exercises non-empty prefixes, so this unbounded-scan failure mode of option 4 was not caught upstream.

What this PR changes

NamedBlobPath.parseS3

String blobNamePrefix = RestUtils.getHeader(args, PREFIX_PARAM, false);
+ if (blobNamePrefix != null && blobNamePrefix.isEmpty()) {
+   blobNamePrefix = null;
+ }

S3-only. Non-S3 callers (which use isListRequest = blobNamePrefix != null to detect a list request) are unchanged so they don't change list-request detection semantics.

MySqlNamedBlobDb.LIST_ALL_QUERY → LIST_ALL_SQL

Convert from static final constant to an instance field initialized via getListAllSQLStatement(config), mirroring the existing LIST_WITH_PREFIX_SQL pattern:

Option	LIST_ALL SQL shape	Semantic
2 or 3	Existing INNER-JOIN + MAX-grouped subquery (preserved verbatim)	If latest version is soft-deleted, second-latest non-deleted surfaces
4	MAX(version) OVER (PARTITION BY blob_name) over single PK range scan, deleted_ts on OUTER select	If latest version is soft-deleted, blob is hidden entirely (matches LIST_WITH_PREFIX_SQL option 4)

For option 2/3 deployments: byte-identical SQL to master. For option 4 deployments: get the window-function plan for both with-prefix and no-prefix paths, and a unified hide-latest-deleted semantic across the two paths.

InMemNamedBlobDb

Handle null prefix to match MySqlNamedBlobDb's LIST_ALL_QUERY path. Previously the in-memory test impl assumed non-null prefix and would NPE in TreeMap.tailMap(null) and entry.getKey().startsWith(null). With parseS3 now normalizing empty to null, the in-memory impl needs the same null-tolerant branching.

Compatibility

• Default listNamedBlobsSQLOption remains 2.
• Window function requires MySQL 8.0+ or TiDB. Already enforced by MySqlNamedBlobDbConfig for option 4 LIST_WITH_PREFIX; no new constraint here.
• The empty-prefix normalization is unconditional but only affects S3 callers that explicitly send prefix= empty.
• Option 4 deployments inherit a behavior change on the no-prefix LIST path: latest-deleted blobs now hide instead of surfacing the second-latest. This unifies the semantic between with-prefix and no-prefix LIST under option 4. Option 2/3 deployments are unaffected.

Testing Done

• ./gradlew :ambry-api:compileJava :ambry-named-mysql:compileJava :ambry-named-mysql:compileTestJava :ambry-named-mysql:compileIntTestJava → BUILD SUCCESSFUL.
• ./gradlew :ambry-api:test --tests com.github.ambry.frontend.NamedBlobPathTest → 10/10 pass, including 3 new cases:
  • testParseS3EmptyPrefixCollapsesToNull
  • testParseS3OmittedPrefixIsNull
  • testParseS3NonEmptyPrefixPreserved
• ./gradlew :ambry-named-mysql:test → all unit tests pass.
• ./gradlew :ambry-frontend:test --tests S3ListHandlerTest → 7/7 pass.
• ./gradlew :ambry-frontend:intTest --tests S3IntegrationTest → 2/2 pass locally.
• The full integration matrix MySqlNamedBlobDbListOperationIntegrationTest already parameterizes over listSqlOption ∈ {2, 3, 4} and will exercise the new LIST_ALL_SQL variants on CI.

Durability risk

Read-only LIST routing change. No write path, no schema, no atomicity, no callback semantics, no resource cleanup. No durability concern.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 46.15385% with 7 lines in your changes missing coverage. Please review.
✅ Project coverage is 51.10%. Comparing base (52ba813) to head (9d18396).
⚠️ Report is 394 commits behind head on master.

Files with missing lines	Patch %	Lines
...ava/com/github/ambry/commons/InMemNamedBlobDb.java	40.00%	1 Missing and 2 partials ⚠️
.../java/com/github/ambry/frontend/NamedBlobPath.java	0.00%	1 Missing and 1 partial ⚠️
.../java/com/github/ambry/named/MySqlNamedBlobDb.java	66.66%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@              Coverage Diff              @@
##             master    #3265       +/-   ##
=============================================
- Coverage     64.24%   51.10%   -13.15%     
+ Complexity    10398     8700     -1698     
=============================================
  Files           840      936       +96     
  Lines         71755    79921     +8166     
  Branches       8611     9579      +968     
=============================================
- Hits          46099    40842     -5257     
- Misses        23004    35684    +12680     
- Partials       2652     3395      +743     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[satishkotha]

do you also want to trim before checking empty?

[zichengl]

considered the trim it but opted not to.

AWS S3 semantics treat prefix=" " as a real filter on names starting with a space rather than as equivalent to no prefix,
and trimming would silently change behavior for any client that legitimately wants to list whitespace-leading blobs.

it also seems like in current Ambry System we allow one empty space but not 2 consecutive spaces
https://github.com/linkedin/ambry/blob/755e2c5450a22f891671918909b733860e77d0cf/am[…]/java/com/github/ambry/frontend/FrontendRestRequestService.java