fix(onboarding): atomic DB append for step completion (deferred bug L3)

markOnboardingStepDone read space_members.onboarding_progress, appended the step
id in JS, and wrote the whole jsonb object back — a lost-update race where two
near-simultaneous completions (double-submit, or two steps ticked quickly) both
read the same base array and the second write clobbered the first, silently
dropping a step.

Replace it with mark_onboarding_step_done (migration 055): a single atomic
UPDATE that dedup-appends the step id into onboarding_progress.completed_step_ids
under the row lock — no update can be lost (READ COMMITTED re-evaluates the jsonb
expression against the latest tuple via EvalPlanQual), other keys are preserved,
and re-appending an existing id is a no-op. Ownership-guarded by auth.uid() so a
caller can only complete steps on their own membership row; the no-JWT service
path is trusted and scoped by p_member_id (mirrors class_signup_tx et al.).

integration/onboarding-progress.test.ts covers dedup/idempotency, key
preservation, the auth.uid ownership gate, and the point of the fix: 12
concurrent completions all survive.

Migration 055 folded into schema.sql; types/database.ts regenerated.
typecheck 0 · lint 0 errors · 645 unit · 58 integration · build ok
This commit is contained in:
Moheeb Zara
2026-08-04 22:36:12 -07:00
parent b2b463491a
commit 4aa6646499
5 changed files with 221 additions and 14 deletions
+110
View File
@@ -0,0 +1,110 @@
// Validates migration 055 (mark_onboarding_step_done) against a real Postgres:
// the atomic dedup-append into space_members.onboarding_progress that replaces
// the old read-modify-write in markOnboardingStepDone (deferred bug L3). The
// central case is CONCURRENCY: many simultaneous completions must all survive —
// the previous SELECT-then-UPDATE could silently drop steps under a lost update.
import { describe, it, expect, beforeAll, afterAll } from 'vitest'
import { promisify } from 'node:util'
import { execFile } from 'node:child_process'
import {
dbReady,
DB_URL,
rows,
execService,
execAsUser,
seedSpace,
seedMember,
dropSpace,
uuid,
} from './db'
const execFileP = promisify(execFile)
const d = describe.skipIf(!dbReady)
d('mark_onboarding_step_done (migration 055)', () => {
let spaceId: string
beforeAll(() => {
spaceId = seedSpace().spaceId
})
afterAll(() => dropSpace(spaceId))
// Completed step ids as the app stores/reads them: onboarding_progress
// -> completed_step_ids (a jsonb string array), or [] when unset.
function completedSteps(memberId: string): string[] {
const r = rows<{ ids: string[] | null }>(
`select (onboarding_progress -> 'completed_step_ids') as ids
from space_members where id = '${memberId}'`,
)
return (r[0]?.ids as string[] | null) ?? []
}
const markService = (memberId: string, stepId: string) =>
execService(`select mark_onboarding_step_done('${memberId}', '${stepId}')`)
it('dedup-appends and is idempotent', () => {
const { memberId } = seedMember(spaceId)
const a = uuid()
const b = uuid()
expect(markService(memberId, a).ok).toBe(true)
expect(markService(memberId, b).ok).toBe(true)
expect(markService(memberId, a).ok).toBe(true) // repeat -> no-op
const ids = completedSteps(memberId)
expect([...ids].sort()).toEqual([a, b].sort())
expect(ids).toHaveLength(2) // 'a' not duplicated
})
it('preserves other keys already in onboarding_progress', () => {
const { memberId } = seedMember(spaceId)
execService(
`update space_members set onboarding_progress = '{"foo":"bar"}'::jsonb where id = '${memberId}'`,
)
const a = uuid()
markService(memberId, a)
const r = rows<{ p: { foo?: string; completed_step_ids?: string[] } }>(
`select onboarding_progress as p from space_members where id = '${memberId}'`,
)
expect(r[0].p.foo).toBe('bar')
expect(r[0].p.completed_step_ids).toEqual([a])
})
it('concurrent completions never lose an update (the L3 fix)', async () => {
const { memberId } = seedMember(spaceId)
const steps = Array.from({ length: 12 }, () => uuid())
// Fire every append at once through separate psql connections. The old
// read-modify-write would have interleaved SELECTs and dropped steps; the
// single-statement atomic UPDATE serializes on the row lock so all survive.
await Promise.all(
steps.map(s =>
execFileP('psql', [
DB_URL,
'-v',
'ON_ERROR_STOP=1',
'-tAc',
`select mark_onboarding_step_done('${memberId}', '${s}')`,
]),
),
)
const ids = completedSteps(memberId)
expect([...ids].sort()).toEqual([...steps].sort()) // all 12 present, none lost
})
it('ownership: with a JWT a member can only complete steps on their OWN row', () => {
const owner = seedMember(spaceId)
const other = seedMember(spaceId)
const step = uuid()
// `other`, acting as a logged-in user, aims the RPC at `owner`'s row.
// auth.uid() != owner.user_id -> the WHERE matches nothing -> no-op.
execAsUser(other.userId, `select mark_onboarding_step_done('${owner.memberId}', '${step}')`)
expect(completedSteps(owner.memberId)).toEqual([])
// The owner completing their own step works.
execAsUser(owner.userId, `select mark_onboarding_step_done('${owner.memberId}', '${step}')`)
expect(completedSteps(owner.memberId)).toEqual([step])
})
})
+9 -14
View File
@@ -130,20 +130,15 @@ export async function markOnboardingStepDone(stepId: string) {
if (!auth.ok) return { error: auth.error }
const { member } = auth
const { data: row } = await supabase
.from('space_members')
.select('onboarding_progress')
.eq('id', member.id)
.single()
const progress = (row?.onboarding_progress ?? {}) as { completed_step_ids?: string[] }
const done = new Set(progress.completed_step_ids ?? [])
done.add(v.data)
const { error } = await supabase
.from('space_members')
.update({ onboarding_progress: { ...progress, completed_step_ids: Array.from(done) } })
.eq('id', member.id)
// Atomic dedup-append in the DB (migration 055). The old read-modify-write on
// onboarding_progress had a lost-update race: two near-simultaneous
// completions both read the same base array and the second write clobbered
// the first, silently dropping a step. The RPC pins the write to the caller's
// own row via auth.uid().
const { error } = await supabase.rpc('mark_onboarding_step_done', {
p_member_id: member.id,
p_step_id: v.data,
})
if (error) return { error: error.message }
return { success: true as const }
@@ -0,0 +1,63 @@
-- =============================================================================
-- 055: atomic, concurrency-safe onboarding step completion
-- =============================================================================
-- markOnboardingStepDone previously did a read-modify-write on the
-- space_members.onboarding_progress jsonb column: SELECT the object, add the
-- step id to completed_step_ids in JS, then UPDATE the whole object back. Two
-- near-simultaneous completions (a double-submit, or two steps ticked in quick
-- succession) both read the same base array and the second write clobbers the
-- first -- a lost update that silently drops a completed step (deferred bug L3).
--
-- mark_onboarding_step_done replaces that with a single atomic UPDATE that
-- dedup-appends the step id into onboarding_progress.completed_step_ids. Because
-- it is one statement, the row lock serializes concurrent callers and Postgres
-- re-evaluates the jsonb expression against the latest committed tuple
-- (EvalPlanQual) under READ COMMITTED -- so no update is ever lost, and other
-- keys in the object are preserved. Appending an already-present id is a no-op
-- (idempotent), matching the old Set semantics.
--
-- Ownership: the id-append is harmless, but a caller must still only be able to
-- touch their OWN membership row. When invoked with a JWT (auth.uid() set, the
-- normal RLS/user client path) the WHERE clause pins the row to
-- user_id = auth.uid(), so passing another member's id updates nothing. The
-- service/no-JWT path (auth.uid() IS NULL) is trusted and scoped by p_member_id,
-- mirroring the existing atomic RPCs (class_signup_tx et al.).
--
-- Idempotent migration: CREATE OR REPLACE. Apply as-is (Supabase SQL editor /
-- psql); re-runs are no-ops.
-- =============================================================================
CREATE OR REPLACE FUNCTION public.mark_onboarding_step_done(
p_member_id uuid, p_step_id uuid
) RETURNS jsonb
LANGUAGE plpgsql SECURITY DEFINER SET search_path = public AS $$
DECLARE
v_completed jsonb;
BEGIN
UPDATE public.space_members m
SET onboarding_progress = jsonb_set(
coalesce(m.onboarding_progress, '{}'::jsonb),
'{completed_step_ids}',
CASE
WHEN coalesce(m.onboarding_progress -> 'completed_step_ids', '[]'::jsonb)
@> to_jsonb(p_step_id::text)
THEN m.onboarding_progress -> 'completed_step_ids'
ELSE coalesce(m.onboarding_progress -> 'completed_step_ids', '[]'::jsonb)
|| to_jsonb(p_step_id::text)
END
)
WHERE m.id = p_member_id
AND (auth.uid() IS NULL OR m.user_id = auth.uid())
RETURNING m.onboarding_progress -> 'completed_step_ids' INTO v_completed;
-- NULL when no row matched (not the caller's membership); the resulting
-- completed_step_ids array otherwise.
RETURN v_completed;
END;
$$;
-- Lock down default PUBLIC execute. The user (RLS) client calls this as the
-- authenticated role; the ownership guard above keeps it to the caller's own
-- row. service_role (the admin client) may also call it for the trusted path.
REVOKE EXECUTE ON FUNCTION public.mark_onboarding_step_done(uuid, uuid) FROM PUBLIC;
GRANT EXECUTE ON FUNCTION public.mark_onboarding_step_done(uuid, uuid) TO authenticated, service_role;
+35
View File
@@ -1655,6 +1655,41 @@ CREATE INDEX IF NOT EXISTS idx_onboarding_steps_space ON public.space_onboarding
ALTER TABLE public.space_members ADD COLUMN IF NOT EXISTS onboarding_completed_at timestamptz;
ALTER TABLE public.space_members ADD COLUMN IF NOT EXISTS onboarding_progress jsonb NOT NULL DEFAULT '{}';
-- Atomic, concurrency-safe step completion (migration 055). Single-statement
-- dedup-append into onboarding_progress.completed_step_ids so two concurrent
-- completions can't lose an update (the old read-modify-write in
-- markOnboardingStepDone could). Ownership-guarded: with a JWT it only touches
-- the caller's own row; the no-JWT service path is trusted and scoped by
-- p_member_id.
CREATE OR REPLACE FUNCTION public.mark_onboarding_step_done(
p_member_id uuid, p_step_id uuid
) RETURNS jsonb
LANGUAGE plpgsql SECURITY DEFINER SET search_path = public AS $$
DECLARE
v_completed jsonb;
BEGIN
UPDATE public.space_members m
SET onboarding_progress = jsonb_set(
coalesce(m.onboarding_progress, '{}'::jsonb),
'{completed_step_ids}',
CASE
WHEN coalesce(m.onboarding_progress -> 'completed_step_ids', '[]'::jsonb)
@> to_jsonb(p_step_id::text)
THEN m.onboarding_progress -> 'completed_step_ids'
ELSE coalesce(m.onboarding_progress -> 'completed_step_ids', '[]'::jsonb)
|| to_jsonb(p_step_id::text)
END
)
WHERE m.id = p_member_id
AND (auth.uid() IS NULL OR m.user_id = auth.uid())
RETURNING m.onboarding_progress -> 'completed_step_ids' INTO v_completed;
RETURN v_completed;
END;
$$;
REVOKE EXECUTE ON FUNCTION public.mark_onboarding_step_done(uuid, uuid) FROM PUBLIC;
GRANT EXECUTE ON FUNCTION public.mark_onboarding_step_done(uuid, uuid) TO authenticated, service_role;
CREATE OR REPLACE FUNCTION public.seed_default_onboarding_steps()
RETURNS trigger LANGUAGE plpgsql SECURITY DEFINER SET search_path = public
AS $$
+4
View File
@@ -3462,6 +3462,10 @@ export type Database = {
}
expire_proposals: { Args: never; Returns: number }
get_user_space_ids: { Args: { uid: string }; Returns: string[] }
mark_onboarding_step_done: {
Args: { p_member_id: string; p_step_id: string }
Returns: Json
}
members_with_permission: {
Args: { perm: string; sid: string }
Returns: {