- New internal/repository/logstore abstraction: exported domain interfaces
(AccessLogStore/ObservabilityStore/UserAccessLogStore/StatusStore),
config-driven provider (Active/Build/Migrating/SetConfigReader), GORM
implementation for PostgreSQL/SQLite (incl. hourly rollups computed in
real time, migration listers, PG partition maintenance), and a ClickHouse
wrapper preserving the native batch path; repository facade delegates to
logstore; import-lint test enforces apps never import analyticsrepo.
- ClickHouse is now optional: the log DB is either the main DB (postgres
when database.enabled, else sqlite) or clickhouse; boot validation +
first-run seed; log_database / log_db_migration are protected keys.
- New user task 切换日志数据库 (of_log_db_switch): freeze log writes,
drain batch writers, copy all 6 raw log tables by id (preserving IDs)
with target-partition pre-creation for PG, flip log_database on success,
clear the freeze flag on failure.
- Per-store retention (log_retention_days_*) with expiry cleanup folded
into the daily system_cleanup task; legacy database_auto_cleanup_* and
of_database_auto_cleanup decommissioned.
- goose migrations: 6 log tables in PG (2 monthly-partitioned) + SQLite,
retention config seeds, schedule cleanup; GET
/api/v1/admin/status/log-database endpoint; frontend retention settings,
switch-task UI and status badge; changelog and docs updated.
docs(plan): log database decoupling implementation plan
docs(design): log database decoupling design (ClickHouse optional)
Prefer capacity/openresty rollups only when they cover the 24h window;
otherwise merge per hour so raw fills pre-MV gaps and rollup wins on
overlap. Add a one-time ANTI JOIN backfill migration for the last 30 days.
Materialized capacity/openresty hourly tables only hold data after the MV
exists. Preferring any non-empty rollup hid full raw history and left 24h
charts with only recent hours. Use rollup only when its earliest bucket
covers the query window start.
Node and dashboard 24h capacity, network, and disk IO charts only used the
latest limited raw snapshots (120/500 rows), so historical hour buckets stayed
empty. Prefer ClickHouse hourly aggregates with counter deltas, and fall back
to raw snapshots when aggregation is unavailable.
ryan