Convert OpenSearch's index-stats merge wait to Solr's wait-for-merges

[serhiy-bzhezytskyy]

Description

OpenSearch Benchmark has no operation type for waiting on merges — workloads poll index-stats until _all.total.merges.current reaches zero. The converter passed that through as "operation-type": "index-stats", which no runner is registered for, so the converted workload aborted with No runner available for operation type [index-stats].

Solr Orbit already has the operation: wait-for-merges, registered in the initial port (57387ed4), and the hand-written Solr nyc_taxis already uses it. Its runner does the polling itself, so the whole idiom converts to it, and the condition / retry-until-success keys become the runner's retry-wait-period and max-wait-seconds. Nothing new is introduced here — the converter just stops emitting an operation type that has no runner.

The condition path is what identifies the idiom, so an index-stats operation whose condition polls something else is left alone. No upstream workload has one, so that guard is defensive; a test pins it.

Both nyc_taxis references live in fragments whose Jinja2 directives keep them from parsing as JSON, which the converter copies verbatim, so the same conversion is applied to their text.

90 lines of code, 164 of tests, 1 of docs.

Issues Resolved

Fixes #79

Testing

• New functionality includes testing

Before / after, same converted workload, one task. Upstream nyc_taxis converted on main at f088380 and on this branch, then:

solr-orbit run --workload-path <converted>/nyc_taxis --pipeline=benchmark-only \
  --target-hosts=localhost:8983 --include-tasks=wait-until-merges-finish

	main	this branch
exit code	64	0
	No runner available for operation type [index-stats]	Using user-provided operation type [wait-for-merges] for operation [wait-until-merges-finish]

The task is include-in-reporting: false, so ✅ SUCCESS on its own proves nothing about it. Reading ~/.solr-orbit/logs/benchmark.log, the schedule is built ([0] warmup iterations and [1] iterations) and on a copy with reporting turned on the task reports 73.207 ms service time at a 0% error rate.

The generated operation now matches the hand-written Solr nyc_taxis in solr-orbit-workloads byte for byte, key order included:

{
  "name": "wait-until-merges-finish",
  "operation-type": "wait-for-merges",
  "retry-wait-period": 2.0,
  "max-wait-seconds": 600,
  "include-in-reporting": false
}

Reach and false positives. Converting every upstream workload that contains the idiom (clickbench, geonames, geoshape, http_logs, neural_search, noaa_semantic_search, nyc_taxis, pmc, treccovid_semantic_search) and counting "index-stats" in the output:

main         17 matches across 14 files
this branch   1 match  (geonames/test_procedures/default.json)

All 16 that changed are operation definitions, and every one is this idiom. The remaining match is not a definition — it is a schedule reference, "operation": "index-stats", to OpenSearch Benchmark's built-in operation of that name, which geonames never defines. This change does not address that, and geonames/test_procedures/default.json comes out byte-identical to what main produces.

Diffing the two trees whole: apart from the conversion timestamp in CONVERTED.md, exactly 13 files differ, 16 hunks in total, and every hunk is this rewrite and nothing else.

Indentation. The text path re-serialises the object, so the step is inferred from the matched object rather than fixed: geoshape's fragments are 2-space and nyc_taxis's are 4-space, and a test covers each. Without that, the rewritten keys landed 14 spaces deep inside geoshape.

pytest tests/ → 1118 passed, 5 skipped. ruff clean.

Not covered. Two source occurrences are in nested directories under test_procedures/, which the converter does not walk, so those files never reach the output at all. And geonames' bare "operation": "index-stats" reference is untouched. Both are separate defects; I will report them separately.

---

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.