mongodb: fix the queries and the file the driver reads

[alexey-milovidov]

Found while running the ClickBench MongoDB queries against the MongoDB compatibility layer of ClickHouse/ClickHouse#68493. Four independent defects, all in mongodb/.

benchmark.sh reads a file that does not exist

export BENCH_QUERIES_FILE="queries.txt"

There is no mongodb/queries.txt; the file is queries.sql, and there never was a queries.txt in this directory. bench_run_query reads $BENCH_QUERIES_FILE, so the driver got no queries at all.

Q20 to Q23 lost their predicate

queries.sql was derived from queries.js with EJSON.stringify, which serializes a JavaScript regular expression literal as an empty document. So

{ $match: { URL: /google/ } }

was committed as

{"$match":{"URL":{}}}

which matches nothing — {} is an equality test against the empty document. The four affected queries are Q20, Q21, Q22 and Q23; they now carry the Extended JSON $regularExpression that EJSON.parse in query turns back into a regular expression. On 33.7M rows of hits, Q20 goes from an empty result to the 3979 rows the SQL returns.

Q37 and Q28 did not implement their SQL

• Q37 — the SQL is ... AND DontCountHits = 0 AND IsRefresh = 0 AND Title <> '' GROUP BY Title, the pipeline filtered URL: {$ne: ""}. It is the $match of Q36 with only the $group changed.
• Q28 — the SQL selects k, AVG(length(Referer)), COUNT(*), MIN(Referer); the pipeline left out MIN(Referer). Its regular expression was also (?:www.)? where the SQL has (?:www\.)?, so an unescaped . matched any character.

The dates are not valid Extended JSON

{"$date": "2013-07-01"} is a bare calendar day. mongosh accepts it because it ends up in new Date(...), but Extended JSON asks for a full instant and other drivers reject it outright — pymongo's json_util raises time data '2013-07' does not match format '%Y-%m-%dT%H:%M:%S', which makes the 7 date-filtered queries unusable outside mongosh. They are {"$date": "2013-07-01T00:00:00Z"} now, the same instant.

queries.js is removed

queries.sql is the file the benchmark driver and the playground read. Keeping the same 43 pipelines a second time in queries.js and copying them across by hand is what dropped the regular expressions in the first place, so the copy is gone and README.md points at queries.sql instead.

Verification

Every line of the new queries.sql was run against a MongoDB wire protocol endpoint over 33.7M rows of hits and compared with clickhouse/queries.sql on the same table. Q20, Q22, Q23, Q28 and Q37 now return exactly the values the SQL does, where before they returned nothing or the wrong ones. The queries that still differ are the ones that are not deterministic on this data — they LIMIT without a total order or cut through a run of ties, and their own SQL result changes between two identical runs — plus Q3, where AVG(UserID) overflows in the SQL and the pipeline's $toDecimal does not, and Q26, where the pipeline projects EventTime on purpose so that the following $sort can see it.

The published results are now stale

Q20, Q21, Q22, Q23, Q28 and Q37 measured either nothing or the wrong work, so mongodb/results/ should be re-run. I do not have a MongoDB installation to do that on the reference hardware.

[alexey-milovidov]

We will merge without rerun. MongoDB is too slow.