Files
cairnobs/metadata
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
..
2026-08-13 17:29:38 -07:00

metadata

PostgreSQL schema and migration tooling for Sentry's control-plane config: dashboards, alert rules, and everything else that isn't log data. See /docs/phase-3-dashboard-design.md and /docs/phase-3-alerting-design.md for why this is a separate database from /storage (ClickHouse) rather than new ClickHouse tables — short version: dashboards and alert state need real row-level locking and transactional read-modify-write, which ClickHouse's MergeTree family doesn't provide.

Schema

Seven tables across three features, one shared database (sentry_metadata):

  • dashboards, dashboard_panels — owned by /api (api/internal/dashboards)
  • notification_targets, alert_rules, alert_state, delivery_log — owned by /alerting
  • audit_log — owned by enterprise/internal/audit (Phase 4). Unlike every other table here, this one is not written through the shared sentry role/pool — see "The audit_writer role" below.

"Owned" here is a documentation convention, not a technical boundary — both services connect to the same Postgres instance/database, each with its own hand-written SQL for the tables it's responsible for. Nothing is shared across service internal/ trees for this, matching the existing repo convention that only /proto is shared code (and even that isn't shared logic, just generated bindings).

The audit_writer role: a second, more restricted credential

audit_log is append-only by design (see /docs/phase-4-isolation-design.md's audit-logging section) — a compliance requirement, not just a convention, so it's backed by two independent defenses, both verified against a live Postgres, not just written:

  1. A dedicated audit_writer Postgres role (migrations/0012-0014) with only INSERT/SELECT grants on audit_log — no UPDATE/DELETE/TRUNCATE, ever. enterprise/internal/audit.Store connects using this role's credentials via its own pgxpool.Pool, never the shared sentry pool api/alerting's other stores use — reusing the shared pool for audit writes would give audit_log's application-level credential the same UPDATE/DELETE grants every other metadata table has, silently defeating the whole point.
  2. A BEFORE UPDATE OR DELETE trigger (migrations/0015-0016) that rejects the operation for any role, including the table owner (sentry) — confirmed live: even sentry needs to explicitly ALTER TABLE audit_log DISABLE TRIGGER audit_log_immutable (a privileged, distinct-from-normal-access operation) before it can modify a row. This is redundant defense-in-depth independent of the grant, protecting against a future migration accidentally re-granting UPDATE to audit_writer.

AUDIT_WRITER_PASSWORD (default audit-writer-dev-only, matching every other dev-only credential in this repo) sets the role's password at creation time via psql -v audit_writer_password=... substitution in migrate.shnot hardcoded in the migration SQL file itself. One real gotcha found while building this: psql's :'var' substitution does not apply inside a DO $$ ... $$ dollar-quoted block (by design, so client-side substitution can't corrupt a function/procedure body) — the role-creation migration is a plain CREATE ROLE, not wrapped in an IF NOT EXISTS check, relying on schema_migrations tracking for idempotency instead (the same pattern Phase 1's non-idempotent ALTER TABLE ... ADD COLUMN migration in /storage already used).

Migration tooling: mirrors /storage/migrate.sh, not a framework

Same reasoning as /storage/README.md: pulling in golang-migrate for what's currently six CREATE TABLE statements is premature machinery. migrate.sh applies migrations/*.sql in filename order over psql, tracking what's applied in a schema_migrations table, one DDL object per file (kept for repo-wide consistency of what a migration "version" means, even though Postgres itself supports multi-statement transactions unlike ClickHouse's HTTP interface).

Running

docker compose up -d                                   # starts a standalone Postgres for local work
POSTGRES_PASSWORD=sentry-dev-only ./migrate.sh          # applies migrations/*.sql

Environment variables migrate.sh reads (all optional except POSTGRES_PASSWORD, matching the root docker-compose.yml's metadata-postgres service):

Var Default
POSTGRES_HOST localhost
POSTGRES_PORT 5432
POSTGRES_USER sentry
POSTGRES_PASSWORD (empty — must be set)
POSTGRES_DATABASE sentry_metadata
AUDIT_WRITER_PASSWORD audit-writer-dev-only

The database itself isn't created by migrate.sh — the postgres:16-alpine image auto-creates POSTGRES_DB on first startup, unlike ClickHouse where migrate.sh has to issue CREATE DATABASE IF NOT EXISTS itself.

There's also a Dockerfile (bash + the postgresql16-client package baked in, migrations/ copied in at build time) used by the root-level docker-compose.yml as a one-shot init service (metadata-migrate) — no runtime package install, no host volume mount needed.

Adding a migration

Add migrations/000N_description.sql with the next sequential number and a single DDL statement. migrate.sh picks it up automatically — no registration step.