mirror of
https://github.com/block/buzz.git
synced 2026-11-11 20:01:09 +00:00
Add node C of the workflow-delivery DAG: the complete dormant DB
transition system for workflow messages addressed to managed agents,
built on B's canonical delivery protocol vocabulary.
Migration 0035 creates workflow_agent_deliveries: one canonical binding
per (community, run, step, target) with a pending -> claimed ->
finished | failed status enum, a monotonic lease_generation fence,
decomposed WorkflowDeliveryCause identity columns guarded by a CHECK
that makes an ambiguous or absent cause unrepresentable, cascading FKs
to workflows/workflow_runs/events, and the community write fence. The
desired-state schema.sql mirrors it inline.
The buzz-db module persists B's WorkflowDeliveryBinding verbatim (no
duplicate tuple spelling) and owns only the lifecycle around it:
- lock_workflow_agent_delivery_identity + commit_workflow_agent_deliveries:
the ONLY insert path, atomic per step, idempotent across producer
retries via ON CONFLICT DO NOTHING.
- claim_workflow_agent_delivery: FOR UPDATE SKIP LOCKED, scoped to the
authenticated target and community; an optional expected binding must
match every supplied field (workflow/run/step/definition/message and
the full decomposed cause) so a forged or stale wake is a miss, never
an alternate authority path. The winning row bumps its fence token.
- renew / finish: fenced by lease_generation; a reaped or superseded
holder matches zero rows and fails closed. finish is terminal-once and
reconciles uncertain completion to a single terminal idempotently.
- reap_expired_workflow_agent_deliveries: fleet-wide, filtered through
community_write_allowed exactly like the scheduler prune scan, so a
quiescing/fenced/deleted tenant is skipped before its write fence can
abort healthy tenants in the same statement.
No producer, runtime, API, or ACP reachability: dormant by contract.
Fresh-Postgres invariant matrix (#[ignore]) proves create idempotency,
community/target claim isolation, mismatched-binding rejection (incl.
cause- and workflow-only disagreement), lease fencing by generation,
reap + prior-holder fencing, terminal-once/idempotent reconciliation,
the fleet-wide lifecycle fence, run-deletion cascade, and commit
rollback. migration.rs asserts len 35, the 0035 contract, and fence
parity.
Co-authored-by: Mongo <89eda16aff8223982a0d928ccc850c50a22abc23babd9f890dc8d6d5431876fb@buzz.block.builderlab.xyz>
Signed-off-by: Logan Johnson <loganj@squareup.com>
91 lines
4.5 KiB
SQL
91 lines
4.5 KiB
SQL
-- Durable, target-scoped delivery inbox and complete transition state machine
|
|
-- for workflow messages addressed to managed agents.
|
|
--
|
|
-- This is the DB-layer complement of the zero-I/O `workflow_delivery` protocol
|
|
-- vocabulary in buzz-core. It persists exactly one canonical binding per
|
|
-- (community, run, step, target) and owns the delivery lifecycle:
|
|
--
|
|
-- pending --claim--> claimed --finish--> finished | failed
|
|
-- ^ |
|
|
-- +------ reap --------+ (expired lease reclaimed; prior holder fenced)
|
|
--
|
|
-- Leases are fenced by a monotonic `lease_generation`: every claim/reclaim
|
|
-- bumps it, and renew/finish only advance a row whose generation still matches
|
|
-- the caller's, so a reaped or superseded holder always fails closed. The same
|
|
-- fleet-wide reaper filters candidate rows through `community_write_allowed`,
|
|
-- exactly like the scheduler prune scan, so a quiescing/fenced/deleted tenant
|
|
-- is skipped before its write-fence trigger can abort healthy tenants.
|
|
--
|
|
-- The producer/runtime/API/ACP nodes are intentionally not reachable here.
|
|
|
|
CREATE TYPE workflow_agent_delivery_status AS ENUM (
|
|
'pending', 'claimed', 'finished', 'failed'
|
|
);
|
|
|
|
CREATE TABLE workflow_agent_deliveries (
|
|
community_id UUID NOT NULL REFERENCES communities(id),
|
|
id UUID NOT NULL,
|
|
workflow_id UUID NOT NULL,
|
|
run_id UUID NOT NULL,
|
|
step_id VARCHAR(64) NOT NULL CHECK (length(btrim(step_id)) > 0),
|
|
target_pubkey BYTEA NOT NULL CHECK (octet_length(target_pubkey) = 32),
|
|
definition_event_id BYTEA NOT NULL CHECK (octet_length(definition_event_id) = 32),
|
|
message_event_id BYTEA NOT NULL CHECK (octet_length(message_event_id) = 32),
|
|
message_event_created_at TIMESTAMPTZ NOT NULL,
|
|
-- Canonical trigger authority identity (buzz-core WorkflowDeliveryCause).
|
|
-- Exactly one identity column is populated per row; the CHECK below makes
|
|
-- an ambiguous or absent cause unrepresentable.
|
|
cause_kind TEXT NOT NULL CHECK (cause_kind IN ('event', 'schedule', 'webhook')),
|
|
cause_event_id BYTEA CHECK (cause_event_id IS NULL OR octet_length(cause_event_id) = 32),
|
|
cause_scheduled_for TIMESTAMPTZ,
|
|
cause_webhook_invocation_id UUID,
|
|
status workflow_agent_delivery_status NOT NULL DEFAULT 'pending',
|
|
-- Monotonic fence token. Bumped on every claim and every reap so a stale
|
|
-- holder's renew/finish matches zero rows and fails closed.
|
|
lease_generation BIGINT NOT NULL DEFAULT 0 CHECK (lease_generation >= 0),
|
|
-- Lease expiry for the current claim; NULL unless status = 'claimed'.
|
|
lease_until TIMESTAMPTZ,
|
|
claimed_at TIMESTAMPTZ,
|
|
finished_at TIMESTAMPTZ,
|
|
created_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
|
|
PRIMARY KEY (community_id, id),
|
|
UNIQUE (community_id, run_id, step_id, target_pubkey),
|
|
FOREIGN KEY (community_id, workflow_id) REFERENCES workflows (community_id, id) ON DELETE CASCADE,
|
|
FOREIGN KEY (community_id, run_id) REFERENCES workflow_runs (community_id, id) ON DELETE CASCADE,
|
|
FOREIGN KEY (community_id, message_event_created_at, message_event_id)
|
|
REFERENCES events (community_id, created_at, id) ON DELETE CASCADE,
|
|
-- Exactly one cause identity is present, matching cause_kind. Any other
|
|
-- combination is a malformed authority and cannot be inserted.
|
|
CHECK (
|
|
(cause_kind = 'event'
|
|
AND cause_event_id IS NOT NULL
|
|
AND cause_scheduled_for IS NULL
|
|
AND cause_webhook_invocation_id IS NULL)
|
|
OR (cause_kind = 'schedule'
|
|
AND cause_event_id IS NULL
|
|
AND cause_scheduled_for IS NOT NULL
|
|
AND cause_webhook_invocation_id IS NULL)
|
|
OR (cause_kind = 'webhook'
|
|
AND cause_event_id IS NULL
|
|
AND cause_scheduled_for IS NULL
|
|
AND cause_webhook_invocation_id IS NOT NULL)
|
|
),
|
|
-- A lease exists iff the row is currently claimed.
|
|
CHECK ((status = 'claimed') = (lease_until IS NOT NULL)),
|
|
CHECK ((status = 'claimed') = (claimed_at IS NOT NULL)),
|
|
-- Terminal rows record when they settled; non-terminal rows never do.
|
|
CHECK ((status IN ('finished', 'failed')) = (finished_at IS NOT NULL))
|
|
);
|
|
|
|
-- Oldest-first polling of claimable work, per authenticated target.
|
|
CREATE INDEX idx_workflow_agent_deliveries_pending
|
|
ON workflow_agent_deliveries (community_id, target_pubkey, created_at)
|
|
WHERE status = 'pending';
|
|
|
|
-- Bounded reaper scan: only currently-leased rows can expire.
|
|
CREATE INDEX idx_workflow_agent_deliveries_lease
|
|
ON workflow_agent_deliveries (lease_until)
|
|
WHERE status = 'claimed';
|
|
|
|
SELECT attach_community_write_fence('workflow_agent_deliveries');
|