Randomize range bounds in Solr's string queries

[serhiy-bzhezytskyy]

Description

--randomization-enabled had no effect on a Solr workload: every client sent the one query written in operations/*.json for the whole run, and the log said randomization was on.

extract_fields_and_paths walks params["body"]["query"] as an object tree looking for a range key, which is the OpenSearch query DSL. Solr's JSON DSL states the query as a string, so the walk reached a leaf immediately and returned no fields; with no fields no standard value source is called and set_range never runs.

On main at f088380, the same nyc_taxis range operation in both shapes:

Solr JSON DSL        -> []
OpenSearch query DSL -> [('total_amount', ['bool', 'filter', 'range'])]

get_randomized_values -> total_amount:[5 TO 15}     # the value source returned {"gte": 3, "lte": 7}

This finds range terms in the query string and in each entry of the filter list, since nyc_taxis' distance_amount_facet states its range there, and substitutes only the bounds:

total_amount:[5 TO 15}                          ->  total_amount:[3 TO 7}
["trip_distance:[0 TO 50}"]                     ->  ["trip_distance:[3 TO 7}"]

The brackets the query already states are kept, so an exclusive bound stays exclusive — the same way the object path keeps whichever of lt/lte it found.

Three deliberate details:

A term that leaves a bound open is left alone. * is Solr's spelling of a bound the query does not state, and the object path does not randomize a one-sided range either: check_one_of_each_name_present requires one name from each of gte/gt and lte/lt, so such a field is never collected. total_amount:[5 TO *] is therefore untouched.

The skip is applied in the substitution as well as the extraction. A term visited by the regexp but absent from fields_and_paths would otherwise consume the value meant for the next term and write it to the wrong field: total_amount:[* TO 15} AND trip_distance:[1 TO 9} with one value gave total_amount:[3 TO 7} AND trip_distance:[1 TO 9}.

Bounds are read by name, not with a default. A value source that returns unexpected keys raises KeyError rather than silently leaving the query as written — the same failure the object path gives, and the same silent no-op this PR removes.

Measured

Not only unit tests. 450 randomized queries over the five nyc_taxis operations that register a value source, across six seeds, against Solr 10.0.0 with 300,649 documents: all HTTP 200, and every matched document lands in a facet bucket — sum(bucket counts) == numFound throughout, 0 violations. The only queries that match nothing are trip_distance:[N TO N}, which upstream produces just as often from gte: 5, lt: 5.

I also checked the wider blast radius, since this makes a previously dead path live: an operation with a two-sided range term but no registered value source now fails a run with SystemSetupError from params.generate_standard_values_if_absent. Scanning every operations file in solr-orbit-workloads: 5 range terms, all 5 registered, 0 missing; geonames has none. So no existing workload starts failing.

⚠️ It does make two latent defects in solr-orbit-workloads nyc_taxis/workload.py reachable — random_money_values passes a float to random.randrange, which raises TypeError on every Python this project supports, so the range operation dies immediately. I have those ready and will open them against solr-orbit-workloads alongside this; they should land together with it.

Issues Resolved

Fixes #75

Testing

• New functionality includes testing

Six tests in tests/workload/loader_test.py: range terms found in a query string and in the filter list, the brackets preserved through set_range, the whole get_randomized_values path on a string query, an open bound left alone, a skipped term not taking another term's value, and a mis-keyed value source raising.

Each of the three details above was mutation-checked — reverting it fails exactly the test that covers it:

reverted	test that fails
guard in extract_solr_range_terms	..._leaves_a_bound_open_is_not_randomized: [('total_amount', ('query',))] != []
guard inside replace()	..._does_not_take_another_terms_value: value written to the wrong field
direct indexing → .get(name, old)	..._omits_a_bound_is_not_silently_ignored: KeyError not raised

ruff check clean; full unit suite 1116 passed / 5 skipped. make it was not run.

[serhiy-bzhezytskyy]

For sequencing: the two workloads defects this makes reachable are apache/solr-orbit-workloads#20. That one is inert without this PR, and this PR breaks --randomization-enabled on nyc_taxis without that one, so they want to go together.