mongodb: give every config its own collection and search index (#306)

[fcostaoliveira]

Closes #306.

The bug

MongoDB addressed one fixed bench.vectors / vector_index for every config, and never adopted the per-config namespacing from #151-4 (used today by redis, valkey, dragonfly, kividb, vectorsets).

Under --skip-upload, the shipped 12-config HNSW sweep (M ∈ {16,32,64} × EF_CONSTRUCTION ∈ {100,200,400,800}) therefore produced 12 result files with 12 distinct HNSW labels, all measured against whichever index the first config built — --skip-upload skips configure(), and configure() is the only caller of create_vector_index.

The reuse precondition could not catch it: corpus_row_count is countDocuments({}), which interrogates the write store, while queries are served by the Atlas search index — a separate system that can be missing, stale, or built with different parameters while the count is perfect.

The fix

collection_name and index_name now derive from engine::index_naming::derive_index_name, following the Redis-family convention exactly. MONGODB_COLLECTION / MONGODB_INDEX_NAME remain overrides but are now bases with the sanitized config name appended, with the usual <BASE>_EXACT=1 escape hatch — the same precedent #151-4 set for REDIS_INDEX_NAME.

MongoDB-specific wrinkle: the server refuses a <db>.<collection> namespace over 255 bytes (verified live against mongodb-atlas-local:8.0.17 — 200 bytes accepted, 250 rejected). Over-budget names become <head>~<16 hex FNV-1a-64 of the whole name>, hashed rather than truncated, so two configs sharing a long prefix cannot collapse back onto one collection. FNV rather than DefaultHasher because SipHash is not guaranteed stable across Rust releases and the name must resolve identically from a later build for --skip-upload to find the corpus. The search index name is left unbounded (a 5000-byte one was accepted).

MongoDB also joins the #151-4 startup collision guard. Its destructible namespace is the collection, not the index — configure() drops the collection and the search index lives inside it — so that is what the guard compares, and it calls the engine's own naming function rather than re-deriving, so the guard checks the name production actually uses.

index_naming itself is untouched: a 255-byte namespace ceiling is MongoDB-specific knowledge with no Redis analogue, so the bound lives in the engine. Zero conflict surface with other in-flight branches there.

Tests

Every test was verified RED by mutation — reverting derivation to env_or, replacing the hash with plain truncation, dropping db_name from the budget, and removing "mongodb" from the guard.

Test	Breaks when
sweep_configs_derive_distinct_collections_and_indexes	derivation reverted
shipped_hnsw_sweep_gives_every_config_its_own_collection	reads the real 12-config file; a colliding 13th config fails here
long_config_names_are_bounded_without_colliding	plain truncation — this is the #294 sanitize-collision path, now tested
the_namespace_budget_accounts_for_the_database_name	budget stops subtracting db_name.len()+1
bounding_is_deterministic	hash swapped (pins literal FNV vectors)
bounding_survives_a_budget_smaller_than_the_digest	slice panic at tiny budgets
bounding_cuts_on_a_char_boundary	mid-UTF-8 slice panic (env base is unsanitized)
collection_env_override_is_a_base_not_a_pin	derivation reverted
test_binary_mongodb_skip_upload_against_another_configs_corpus_is_fatal	the headline regression — config B reusing config A's corpus must abort, and A's corpus must survive
test_binary_mongodb_exact_pin_with_two_configs_is_rejected_at_startup	"mongodb" removed from the guard; asserts the error text
test_binary_mongodb_keep_data_keeps_every_configs_corpus	pins the experiment.rs:298-305 claim, which is now true for MongoDB

Verification

make check clean · cargo test --bin vector-db-benchmark 883 pass · cargo test --lib 129 pass · MONGODB_PORT=27018 cargo test --test integration_mongodb 25 pass · overhead_invariants 9 pass · integration_vectorsets exact_pin passes (its asserted substring is preserved).

Reviewer notes — deliberate calls worth challenging

1. MONGODB_COLLECTION / MONGODB_INDEX_NAME semantics change from "exact name" to "base name". Anyone scripting against a fixed collection now gets vectors:<config>; _EXACT=1 restores the old behaviour. This follows the Vertex benchmark reliability: 401 handling, token refresh, --engines-file no-op, Redis M×EFC index collision, search watchdog #151-4 precedent, but it is a user-visible break and no doc outside the engine mentions these vars.
2. The index name is namespaced too, not just the collection. Per-config collections alone fully isolate (an Atlas search index is scoped to its collection), and the guard now rejects the pinned shared-collection case anyway — so this half arguably buys nothing today. It matches the issue text and costs nothing; dropping it is a clean one-line revert if you prefer a smaller blast radius on MONGODB_INDEX_NAME.
3. No identity marker added to corpus_row_count. Per-config naming makes a cross-config --skip-upload count 0 and fail, so a marker would answer the same question one layer later. The dataset dimension of --skip-upload verifies corpus SIZE, not corpus IDENTITY (cardinality is not identity) #279 (one config reused across two same-sized datasets) is deliberately left open — that is identical on all 15 engines and wants one cross-engine mechanism.

Merge order

This conflicts with #305's branch in mongodb_engine.rs and tests/integration_mongodb.rs (both edit the same file in different regions). #305 fixes a higher-severity bug — recall published against a ~1%-indexed index — so land #305 first and I will rebase this.

🤖 Generated with Claude Code

[github-actions]

Docker Build Validation

Platforms tested:

• linux/amd64 (built and tested)
• linux/arm64 (build validated)

Tests performed:

• --help command
• --describe datasets command
• Integration test (random-100 dataset + Redis)
• Multi-platform build validation (amd64 + arm64)

[github-actions]

Docker Build Validation

Platforms tested:

• linux/amd64 (built and tested)
• linux/arm64 (build validated)

Tests performed:

• --help command
• --describe datasets command
• Integration test (random-100 dataset + Redis)
• Multi-platform build validation (amd64 + arm64)