cairnobs/search

search

Tantivy-backed full-text search over log messages. Phase 1's answer to "grep across everything," separate from ClickHouse's structured/ aggregation queries.

Why a separate service, not embedded in ingest

Tantivy is a Rust library with no maintained Go bindings — using it from ingest (Go) would mean cgo-bridging to a compiled Rust cdylib, exactly the fragile FFI complexity CLAUDE.md's "prefer boring, well-understood dependencies... operators need to trust it" principle steers away from. It would also couple ClickHouse-write latency to Tantivy-write latency in the same request path. See /docs/architecture.md for the fuller tradeoff writeup (dual-write vs. second-consumer-group) from Phase 1 planning.

How it fits together

ingest (gRPC front end) --> Redpanda (sentry.logs.raw) --> ingest's ClickHouse-writer consumer --> ClickHouse
                                    \
                                     `--> search's own consumer --> Tantivy index

search reads the same Redpanda topic ingest's ClickHouse-writer consumer reads, as an independent consumer in spirit (own offset tracking, own failure domain — see "Offset tracking" below) even though it's a completely separate process/service. If Tantivy indexing lags or crashes, ClickHouse ingestion is completely unaffected.

api (or, for a multi-tenant deployment, enterprise-api — see /enterprise/README.md) calls search's SearchService.Search gRPC RPC (see /proto/sentry/search/v1/search.proto) to resolve a free-text query into matching record_ids, then joins those back against ClickHouse's logs.record_id column (see /storage/migrations/0002_add_record_id.sql) to get full rows. search only ever returns IDs, never row data — it stays a pure text index, not a second copy of the row.

Per-tenant indices (Phase 4) — read and write

SearchRequest.tenant_id (empty by default) selects which index src/registry.rs's IndexRegistry searches: empty resolves to the single default index every deployment already had; a non-empty value opens (on first use) a dedicated index under TENANTS_INDEX_PATH (default /var/lib/sentry-search/tenants/<tenant_id>, matching deploy/operator's and enterprise/internal/rbacstore's existing path convention). tenant_id is set only by a trusted server-side caller (enterprise/internal/searchclient, from the authenticated request identity) — never a value a browser/client controls.

consumer.rs's Redpanda consumer resolves the same registry now, keyed by each Kafka message's tenant_id header — attached server-side by ingest/internal/grpcserver after validating an agent's per-tenant credential, mirroring exactly how enterprise/cmd/enterprise-ingest routes the ClickHouse write side (see /enterprise/README.md's "Ingest write-routing" section). No import boundary to work around here, unlike Go: IndexRegistry already lives in this AGPL-core binary, so both the read and write paths share one registry directly, with no "second binary" needed. The periodic Tantivy commit (COMMIT_INTERVAL_MS) now commits every tenant index that's actually seen a write, plus the default index, via IndexRegistry::commit_all, not just one index.

The active-tenant gap is closed too, the same way the read side closes it: src/tenants.rs's ActiveTenantTracker polls a new GET /internal/active-tenants endpoint on enterprise-auth (this process has no Postgres access, so unlike enterprise/internal/ searchclient's TenantChecker — a direct rbacstore.TenantIsActive call, since that code runs in enterprise/ — this needed a network call instead), and consumer.rs refuses (logs and skips, never falls back to the default or another tenant's index) any tagged record whose tenant_id isn't in the polled allowlist. Off unless ENTERPRISE_AUTH_URL/ENTERPRISE_AUTH_SERVICE_TOKEN are both set (see Configuration below) — when they aren't, write-routing behaves exactly as it did before this tracker existed, trusting any syntactically-valid tenant_id. When they are, startup blocks on the first fetch succeeding (fail-closed cold start — see tenants.rs's doc comment for why a partial/degraded startup isn't the safer choice, and for the last-known-good behavior periodic refresh failures fall back to). Verified with real HTTP round trips against a hand-rolled TCP test server (no mocking crate needed for one endpoint) — the Bearer token actually sent, the initial-fetch-fails-closed path, and an unreachable server also failing closed. See src/registry.rs's doc comment on resolve for how the mechanism (index lifecycle) and policy (who gets gated) responsibilities are split.

Offset tracking: why this isn't a Kafka consumer group

rskafka (chosen for being pure Rust, no cgo — consistent with why ingest chose segmentio/kafka-go over confluent-kafka-go) is a low-level client: it doesn't implement Kafka's broker-side consumer-group coordination protocol the way kafka-go or librdkafka do. So search tracks its own per-partition offsets in a plain JSON file next to the Tantivy index (offsets.rs), persisted after each fetched batch.

This is deliberately best-effort, not exactly-once: if the process dies between processing a record and persisting its offset, that record gets reprocessed after restart. This is safe because SearchIndex::upsert is delete-then-add on record_id — Tantivy segments are immutable, so this is the standard idiom for updates anyway, and it happens to make reprocessing idempotent for free. Partition count is read from REDPANDA_TOPIC_PARTITIONS (must match what /transport/provision-topics.sh actually created — same kind of documented cross-component contract as the topic name itself), not discovered dynamically.

Query syntax

Whatever Tantivy's own QueryParser supports against the message field: plain terms, "exact phrase" queries, and foo* wildcards. Not documented further here because it's Tantivy's syntax, not Sentry's — see Tantivy's query parser docs for the full grammar. No unified query language yet; that's Phase 2.

Configuration

Environment variables (see src/config.rs):

Var	Default	Purpose
GRPC_LISTEN_ADDR	0.0.0.0:50052	Full socket address — Rust's parser needs one, unlike Go's :PORT shorthand ingest/api use
REDPANDA_BROKERS	localhost:9092	Comma-separated broker list
REDPANDA_TOPIC	sentry.logs.raw	Must match /ingest's topic
REDPANDA_TOPIC_PARTITIONS	6	Must match what /transport/provision-topics.sh created
INDEX_PATH	/var/lib/sentry-search/index	Default (non-tenant) Tantivy index directory
TENANTS_INDEX_PATH	/var/lib/sentry-search/tenants	Per-tenant index directories live under here, one subdirectory per tenant_id (Phase 4)
OFFSETS_PATH	/var/lib/sentry-search/offsets.json	Offset tracking file
COMMIT_INTERVAL_MS	2000	How often buffered writes become searchable
ENTERPRISE_AUTH_URL	(empty)	Enables tenants::ActiveTenantTracker -- empty means write-routing has no active-tenant gate, same as every deployment before Phase 4. Must be set together with ENTERPRISE_AUTH_SERVICE_TOKEN below, or Config::load fails
ENTERPRISE_AUTH_SERVICE_TOKEN	(empty)	RoleService Bearer credential for GET /internal/active-tenants, minted via enterprise-auth -mint-service-token search

Building & testing

cargo build --release
cargo clippy --all-targets -- -D warnings
cargo test

index.rs's tests run against a real (temp-directory) Tantivy index — no external service needed, unlike ClickHouse. They cover the delete-then-add idempotency, phrase queries, result limits, and the before-commit/after-commit visibility boundary. registry.rs's tests are the same shape and cover the actual adversarial claim: real per- tenant indices, real documents, and a search scoped to one tenant returning zero results for another tenant's matching document (tenant_index_is_isolated_from_default_and_other_tenants), plus commit_all_commits_default_and_every_opened_tenant_index proving the write-routing commit path actually makes every tenant's buffered writes searchable, not just the default index's — all of this, unlike almost everything else in Phase 4, was genuinely run in the environment this was built in, not just written. consumer.rs's Kafka/rskafka wiring itself is still not unit-tested — that needs a real Redpanda, same category of gap as /ingest's kafka.Reader/kafka.Writer wiring, and is exercised by the docker-compose end-to-end flow instead — but the pure tenant_id_from_headers header-extraction helper it uses is factored out and unit-tested the same way ingest/consumer's Go equivalent is, including a guard test (test_tenant_id_header_key_matches_go) against the header-key literal drifting from the Go side's. tenants.rs's tests genuinely exercise reqwest against a real (if hand-rolled, dependency-free) TCP server — the actual Authorization: Bearer header construction, JSON response parsing, and both fail-closed paths (a rejected first fetch, an unreachable server), not a fake HTTP client substituted in.

# from the repo root, not search/
docker build -f search/Dockerfile -t sentry-search .