Files
cairnobs/ingest
jcoffey-dev 93c160ec51 Add agent restart lifecycle command
Extends the existing CheckIn RPC with a one-shot AgentCommand
(restart only -- stop/uninstall need real per-platform OS
service-manager integration and stay deliberately out of scope),
delivered at-most-once: cleared the instant it's handed to the agent
in a response, since a restarting agent's process is gone before it
could ever confirm receipt. On restart, the agent flushes whatever's
buffered, aborts its source task, and exits cleanly, relying entirely
on the host's own service manager to bring it back up.

Issuing a command is gated at RoleAdmin (stricter than config
editing's RoleEditor) and logged into the same audit_log table Phase
7's AI interactions use, via a new agent_command event type.

A real bug was found and fixed during live verification: the first
implementation tried to atomically read-and-clear pending_command in
a single INSERT...ON CONFLICT statement using a sibling CTE
referenced only from RETURNING, on the assumption that Postgres
evaluates every part of a WITH query against one pre-statement
snapshot. That's wrong specifically for FOR UPDATE, which always
reads the latest row version including one written earlier in the
same statement -- confirmed empirically (a restart command was
always coming back empty even when genuinely pending, so the agent
never received it). Fixed by splitting into two real, ordered
statements inside one explicit transaction.

See /docs/agent-management-design.md's "Lifecycle commands" section.
2026-08-16 20:30:07 -07:00
..

ingest

Go service sitting between the Rust agent and ClickHouse. Two halves in one binary, selected with --mode:

  • server — mTLS gRPC front end (LogIngest.PushBatch) that agents connect to. Assigns each record a server-side record_id (a UUID, overwriting whatever the agent sent — agents always send it empty) and otherwise forwards records proto-encoded onto Redpanda unchanged. Still kept thin — one field assignment, no real normalization — so agent- facing latency isn't coupled to ClickHouse write performance. record_id has to be assigned exactly once, here, rather than independently by each downstream consumer: Phase 1's Tantivy indexer and the ClickHouse writer both read the same Redpanda messages and need to agree on the same ID for the same record to join search hits back to rows — two consumers generating their own IDs would produce mismatched ones for what's supposed to be the same record. Also (Phase 4): resolves an optional per-tenant Authorization: Bearer <token> credential via internal/grpcserver.TenantResolver (nil by default -- single-tenant behavior unchanged) and attaches the resolved tenant ID to every produced Kafka message as a tenant_id header (consumer.TenantIDHeaderKey) -- see "Multi-tenant write-routing" below.
  • consumer — reads back off Redpanda, normalizes into the ClickHouse row shape (internal/normalize), and batch-writes via the native protocol driver. Commits Redpanda offsets only after a successful ClickHouse write, so a ClickHouse outage causes redelivery on restart rather than data loss. Reads each message's tenant_id header (if any) but this package's own writer (clickhousewriter.Writer, used by cmd/ingest's single-tenant mode) ignores it -- every record still lands in the one shared ClickHouse database regardless of tag. See "Multi-tenant write-routing" below for where the tag actually gets used.
  • all (default) — both, in one process. This is what docker-compose runs. Splitting into two deployments later (e.g. to scale them independently in k8s) is a manifest change, not a code change — see --mode.

Multi-tenant write-routing

This package (AGPL core) only ever writes to one shared ClickHouse database, regardless of any tenant_id tag a message carries -- routing a tagged record into its own tenant's dedicated database is enterprise/internal/chwriter and enterprise/cmd/enterprise-ingest's job (a separate module by architectural convention, not a licensing split -- both are AGPLv3, see /CLAUDE.md's licensing boundary), not this package's. consumer and clickhousewriter live outside internal/ (moved there once enterprise/internal/chwriter needed to import them directly -- Go's compiler-enforced internal/ visibility rule blocks a separate module from importing anything under ingest/internal/..., the same reason several api/internal/... packages moved out earlier in Phase 4) specifically so enterprise/ can reuse this package's own flush loop and ClickHouse batch-insert logic unchanged, rather than reimplementing either. See /enterprise/README.md's "Ingest tenant identity"/write-routing sections for the full story, including what's still not built.

Why Redpanda stays in the path

Confirmed with the project owner during Phase 0 planning: the gRPC front end produces to Redpanda rather than writing ClickHouse directly. This exercises the pinned transport layer from day one and keeps agents from ever needing Kafka credentials — mTLS to ingest is the only network egress an agent has. See /docs/architecture.md.

Dependencies worth knowing about

  • github.com/segmentio/kafka-go — pure Go, no cgo, chosen over franz-go/confluent-kafka-go specifically to keep the distroless build simple (confirmed with the project owner; see git history / PR discussion for the tradeoffs considered).
  • github.com/ClickHouse/clickhouse-go/v2 — official client, native protocol, pure Go (no cgo).
  • golang.org/x/sync/errgroup — used in cmd/ingest/main.go to run the server and consumer halves concurrently and propagate the first error.
  • github.com/google/uuid — was already in the dependency graph transitively (via clickhouse-go); promoted to a direct dependency for record_id generation in internal/grpcserver, so not a new addition to the transitive tree.

Configuration

All via environment variables (see internal/config/config.go for the full list and defaults) — no config file format for Phase 0:

Var Default Purpose
GRPC_LISTEN_ADDR :4317 Agent-facing gRPC listen address
TLS_CERT_FILE / TLS_KEY_FILE /etc/sentry-ingest/server{,-key}.pem ingest's own mTLS identity
TLS_CLIENT_CA_FILE /etc/sentry-ingest/ca.pem CA used to verify agent client certs
REDPANDA_BROKERS localhost:9092 Comma-separated broker list
REDPANDA_TOPIC sentry.logs.raw Must match the topic provisioned in /transport
REDPANDA_CONSUMER_GROUP sentry-ingest Consumer group id
CLICKHOUSE_ADDR localhost:9000 Native protocol port, not HTTP
CLICKHOUSE_DATABASE / _USERNAME / _PASSWORD sentry / default / ``
CONSUMER_BATCH_MAX_SIZE 500 Records per ClickHouse batch insert
CONSUMER_BATCH_FLUSH_INTERVAL_MS 2000 Max time a partial batch waits before flushing
ENTERPRISE_AUTH_URL (empty) Enables internal/grpcserver.TenantResolver -- empty means PushBatch never requires a bearer credential and no tenant_id header is ever attached, same as every Phase 0-3 deployment

Building & testing

go build ./...
go vet ./...
go test ./...

Requires google.golang.org/protobuf/cmd/protoc-gen-go and google.golang.org/grpc/cmd/protoc-gen-go-grpc only if you're regenerating /proto's Go bindings — ingest itself just imports the already-generated github.com/sentry/sentry/proto module (see the replace directive in go.mod, pointing at ../proto).

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

Testing notes

consumer and internal/grpcserver depend on Redpanda and ClickHouse only through small interfaces (reader/chWriter in consumer, batchProducer in grpcserver), so the flush/commit/error-handling logic is unit-tested against fakes — no embedded broker or database needed. This includes the tenant_id tagging/extraction round trip end to end (a fake TenantResolver in grpcserver's tests, a fake Kafka header in consumer's) — real logic, fake transport, no live enterprise-auth or Redpanda needed. What's not covered by these tests: the real kafka.Reader/kafka.Writer wiring and the ClickHouse native-protocol driver itself. Those are only exercised by the docker-compose end-to-end flow described in /docs/phase-0-runbook.md.