multica/server/migrations/101_task_usage_hourly_schem...

134 lines
6.7 KiB
SQL

-- Hourly rollup table for `task_usage`, materialised in **UTC**. Replaces
-- both per-runtime `task_usage_daily` (073, 082) and per-workspace
-- `task_usage_dashboard_daily` (084) as the single source of truth for
-- all token-usage reports. See docs/timezone-architecture-rfc.md §4.
--
-- WHY HOURLY + UTC:
-- The two existing rollups materialise on a `DATE` bucket — one in the
-- runtime's IANA tz, the other in UTC — which forces every report to
-- either accept the materialised tz or scan raw `task_usage`. Hourly
-- UTC buckets are tz-neutral: any viewer-side tz can be applied at
-- query time via `DATE(bucket_hour AT TIME ZONE @tz)` without losing
-- precision and without crossing midnight in the wrong direction.
--
-- WHY ONE TABLE INSTEAD OF TWO:
-- The two existing rollups share the same source rows and the same
-- invalidation surface (atq, task_usage, issue.project_id); maintaining
-- them separately is duplicative. The unified PK carries runtime_id,
-- agent_id, AND project_id, so:
-- * Runtime-detail views filter on runtime_id (covered by
-- idx_..._runtime_time).
-- * Workspace-dashboard views filter on workspace_id + group by
-- agent_id / project_id (covered by the three workspace indexes).
-- * The hour-of-day heatmap groups by EXTRACT(HOUR FROM ... AT TIME
-- ZONE <viewer's tz>) over the same rows — no separate aggregate.
--
-- WHY PROVIDER+MODEL IN THE PK:
-- Per-model breakdowns are a primary read dimension (cost per model,
-- trend per model). Keeping them in the PK keeps the rollup pre-grouped
-- along the same axis the UI uses.
--
-- WHY `UNIQUE NULLS NOT DISTINCT`:
-- `project_id` is nullable — tasks linked to issues without a project,
-- and the quick-create path's "no issue yet" state, both produce
-- no-project usage. PG15's `UNIQUE NULLS NOT DISTINCT` lets ON CONFLICT
-- upsert the no-project bucket the same way it handles a concrete
-- project. (Same pattern as 084.)
CREATE TABLE task_usage_hourly (
bucket_hour TIMESTAMPTZ NOT NULL, -- UTC, truncated to hour boundary
workspace_id UUID NOT NULL,
runtime_id UUID NOT NULL,
agent_id UUID NOT NULL,
project_id UUID, -- nullable; see above
provider TEXT NOT NULL,
model TEXT NOT NULL,
input_tokens BIGINT NOT NULL DEFAULT 0,
output_tokens BIGINT NOT NULL DEFAULT 0,
cache_read_tokens BIGINT NOT NULL DEFAULT 0,
cache_write_tokens BIGINT NOT NULL DEFAULT 0,
task_count BIGINT NOT NULL DEFAULT 0,
event_count BIGINT NOT NULL DEFAULT 0,
updated_at TIMESTAMPTZ NOT NULL DEFAULT now(),
CONSTRAINT uq_task_usage_hourly_key
UNIQUE NULLS NOT DISTINCT
(bucket_hour, workspace_id, runtime_id, agent_id, project_id, provider, model)
);
-- Workspace-wide trend (no other filter): /{slug}/dashboard. The leading
-- workspace_id matches every dashboard query; bucket_hour DESC avoids an
-- extra sort when the report walks "last 7/30/90 days" backwards.
CREATE INDEX idx_task_usage_hourly_workspace_time
ON task_usage_hourly (workspace_id, bucket_hour DESC);
-- Runtime detail page — trend + hour-of-day heatmap on a single runtime.
-- The heatmap groups by `EXTRACT(HOUR FROM bucket_hour AT TIME ZONE
-- <viewer's tz>)` over this range, so we want the rows pre-clustered
-- by runtime.
CREATE INDEX idx_task_usage_hourly_runtime_time
ON task_usage_hourly (runtime_id, bucket_hour DESC);
-- Workspace dashboard "by agent" panel.
CREATE INDEX idx_task_usage_hourly_workspace_agent_time
ON task_usage_hourly (workspace_id, agent_id, bucket_hour DESC);
-- Workspace dashboard "by project" panel. Partial because no-project
-- buckets aggregate into a separate bucket and the panel filters them
-- out; this keeps the index small.
CREATE INDEX idx_task_usage_hourly_workspace_project_time
ON task_usage_hourly (workspace_id, project_id, bucket_hour DESC)
WHERE project_id IS NOT NULL;
-- Single-row state table tracking the rollup worker's watermark. Same
-- shape as 073's `task_usage_rollup_state` and 084's
-- `task_usage_dashboard_rollup_state` — a SMALLINT(1) PK is the easiest
-- way to enforce "exactly one row" without a CHECK trigger.
CREATE TABLE task_usage_hourly_rollup_state (
id SMALLINT PRIMARY KEY DEFAULT 1 CHECK (id = 1),
watermark_at TIMESTAMPTZ NOT NULL DEFAULT '1970-01-01 00:00:00+00',
last_run_started_at TIMESTAMPTZ,
last_run_finished_at TIMESTAMPTZ,
last_run_rows BIGINT NOT NULL DEFAULT 0,
last_error TEXT
);
INSERT INTO task_usage_hourly_rollup_state (id) VALUES (1) ON CONFLICT DO NOTHING;
-- Dirty queue for invalidations the `updated_at` watermark cannot see:
-- * DELETE on `task_usage` (no row left for the watermark to catch).
-- * Cascade DELETE through `agent_task_queue` (task_usage rows gone).
-- * UPDATE of `issue.project_id` — moves the bucket to a new key,
-- OLD bucket needs to shrink, NEW bucket needs to appear.
-- * UPDATE of `agent_task_queue.runtime_id` / `agent_task_queue.issue_id`
-- — same re-attribution problem on different dimensions.
--
-- bucket_hour is computed in UTC at trigger time, so dirty keys match
-- the rollup table byte-for-byte and the window function can UNION the
-- queue into `dirty_keys` without translation.
--
-- TTL: rows in this queue MUST be pruned (see prune_task_usage_hourly_dirty
-- in the rollup-pipeline migration). Without TTL, dense workloads grow the queue
-- unboundedly — every retouched
-- bucket leaves a row behind. The window function deletes rows whose
-- enqueued_at < p_to as part of each tick, which keeps the steady state
-- bounded; the explicit prune is a belt-and-braces guarantee for rows
-- that somehow escape the window (e.g. crash mid-tick).
CREATE TABLE task_usage_hourly_dirty (
bucket_hour TIMESTAMPTZ NOT NULL,
workspace_id UUID NOT NULL,
runtime_id UUID NOT NULL,
agent_id UUID NOT NULL,
project_id UUID,
provider TEXT NOT NULL,
model TEXT NOT NULL,
enqueued_at TIMESTAMPTZ NOT NULL DEFAULT now(),
CONSTRAINT uq_task_usage_hourly_dirty_key
UNIQUE NULLS NOT DISTINCT
(bucket_hour, workspace_id, runtime_id, agent_id, project_id, provider, model)
);
-- The window function drains rows with enqueued_at < p_to; the prune
-- helper (prune_task_usage_hourly_dirty) deletes rows
-- whose enqueued_at falls outside the retention horizon. Both scans
-- use this index.
CREATE INDEX idx_task_usage_hourly_dirty_enqueued_at
ON task_usage_hourly_dirty (enqueued_at);