cd /news/ai-products/we-deleted-10-real-users-with-a-test… · home topics ai-products article
[ARTICLE · art-16327] src=dev.to pub= topic=ai-products verified=true sentiment=↓ negative

We Deleted 10 Real Users with a Test-Cleanup Script — RCA

On May 11, 2026, a test-cleanup script on HoneyChat, a Telegram-native AI companion with approximately 300 daily active users, accidentally deleted 10 real OAuth users from the production PostgreSQL 16 database. The script ran a range-based DELETE command targeting test user IDs between -91111200 and -91111100, but real users had been assigned IDs in that same narrow window due to an unintended interaction between the OAuth ID allocation system and hardcoded test IDs. Three independent design flaws—a sequence advancement mechanism, an upsert pattern that silently overwrote data, and the range-based cleanup—combined to make the deletion possible, with recovery from backup proving effectively impossible due to the 22-hour pg_dump cycle.

read6 min publishedMay 28, 2026

On 2026-05-11, a test-cleanup script on HoneyChat (Telegram-native AI companion, ~3 months in production, ~300 DAU, PostgreSQL 16 + Redis) ran:

DELETE FROM users WHERE id BETWEEN -91111200 AND -91111100;

About ten real OAuth users had IDs in that narrow window. They were now gone. Their users

row, their subscriptions

row, their chat_sessions

/ web_messages

— all gone from Postgres, and recovery from backup was effectively impossible (more on that below).

This is the postmortem and the contract we now run instead. The honest version: the destructive script went to prod on a schema I never verified end-to-end. Three separate design mistakes lined up to make it possible, and not one of them was caught before the script ran on a Tuesday night.

Two signup paths feed the users

table:

Population ID source
Telegram users (most of base) Positive integers — Telegram's own user IDs come in on the message envelope
OAuth users (Google / Discord, web sign-in) Negative integers from a Postgres sequence web_user_id_seq

OAuth IDs were negative on purpose — to keep them out of the positive Telegram-ID space and avoid collisions when a Telegram user later signed in via web. The minter in api/web_auth.py

looked roughly like this:

async def _allocate_negative_user_id(db) -> int:
    for _ in range(5):  # retry on rare UniqueViolation
        new_id = -(await db.fetchval("SELECT nextval('web_user_id_seq')"))
        try:
            await db.execute("INSERT INTO users (id, ...) VALUES (%s, ...)", new_id)
            return new_id
        except UniqueViolation:
            await db.execute(
                "SELECT setval('web_user_id_seq', GREATEST(-MIN(id), currval('web_user_id_seq')))"
                " FROM users"
            )
    raise RuntimeError("could not allocate user id after 5 retries")

The setval(GREATEST(-MIN(id), current))

step is the load-bearing piece you have to keep in mind. It says: whatever the most-negative users.id is right now, my sequence should be at least that far advanced, so I never collide with it again.

For QA I was creating test users by hand with hardcoded negative IDs like -91111101

, -91111102

, … via INSERT ... ON CONFLICT (id) DO UPDATE

. Easy to remember, easy to clean up later by range.

That choice triggered three independent failure modes, each on its own benign, lethal in combination:

web_user_id_seq

to 91 111 101.setval(GREATEST(...))

line above, the very next OAuth signup retry saw the new test row with id = -91111101

, computed -MIN(id) = 91111101

, and advanced its own sequence. From that moment on, all real OAuth signups were drawing IDs in the neighbourhood of -91111111

, -91111112

, … — right inside the window where my test users lived.INSERT ... ON CONFLICT (id) DO UPDATE

.plan

, auth_source

and several other fields instead of erroring.DELETE … WHERE id BETWEEN -91111200 AND -91111100

None of these three behaviors is exotic. The setval(GREATEST(...))

retry pattern is a normal way to handle UniqueViolation on a seeded sequence. ON CONFLICT DO UPDATE

is a normal Postgres upsert. Range-DELETE is a normal cleanup pattern. Each was safe on its own; the interaction of all three was lethal — and I never set up a staging run that would have surfaced the interaction before it touched prod.

A 30-second sanity check on the second insert ("did adding id = -91111101

move web_user_id_seq

? what does the next OAuth signup land on?") would have shown the cascading effect immediately. Nobody — me — ran it. The cleanup script ran nightly for weeks looking healthy because real OAuth signup volume hadn't yet pushed a real ID into the deletion window.

Recovery from Postgres backup was effectively impossible. The chain:

pg_dump

to Storj was about 22 hours old — ON CONFLICT DO UPDATE

had already mutated their plan

and auth_source

columns earlier the same day.users

pages were gone too.What we could salvage came from side channels:

PERSIST

-ed what looked important and reconstructed recent conversations for affected users.chat_sessions

and web_messages

Net: people kept their accounts and most of their recent conversations, but lost web-side scene context older than the Redis window. We comped the affected users. The cost of the incident wasn't the rows — it was the trust dent and the day-and-a-half of recovery work.

ON CONFLICT DO UPDATE

  • range-DELETE) was never verified end-to-end before any of it touched production.INSERT

of id = -91111101

in staging followed by SELECT id FROM users ORDER BY id LIMIT 5

, would have shown the sequence had jumped to the test neighbourhood. Nobody ran it. BETWEEN

query can sweep. An attribute is something a WHERE auth_source = 'test'

query cannot accidentally trip over.INSERT ON CONFLICT (id) DO UPDATE

.INSERT

would have failed loudly and surfaced the collision Any one of these five would have saved us; we had all five wrong.

ALTER TABLE users ADD COLUMN auth_source text NOT NULL DEFAULT 'oauth';
-- backfill: 'telegram' for positive Telegram IDs, 'oauth' for legacy negative,
-- 'test' for known test rows that we then deleted via the new path.
CREATE INDEX users_auth_source_idx ON users(auth_source);
TEST_ID_RANGE = (1_000_000_001, 1_999_999_999)   # high *positive* — out of all real paths

def create_test_user() -> int:
    user_id = _next_test_id()
    db.execute(
        "INSERT INTO users (id, auth_source, ...) VALUES (%s, 'test', ...)",
        (user_id, ...),
    )
    return user_id
php
def cleanup_test_users(dry_run: bool = True) -> int:
    rows = db.fetchall("SELECT id FROM users WHERE auth_source = 'test'")
    if dry_run:
        print(f"Would delete {len(rows)} test users")
        return len(rows)
    db.execute("DELETE FROM users WHERE auth_source = 'test'")
    return len(rows)

The script defaults to dry_run=True

. The CLI flag to actually run it is explicit and shows the count first.

We've also banned, in our engineering doc and in code review: any DELETE … WHERE id BETWEEN …

on the users

table, for any reason; any INSERT … ON CONFLICT (id) DO UPDATE

on users.id

.

We rebuilt the backup story around explicit recovery point objectives. Off-site is Storj (~7 GB total, ~$0.03/month — cost is not the constraint).

Backup tier Cadence Destination RPO
Postgres pg_dump (logical)
Hourly Local disk ≤ 1 h
Postgres pg_dump (logical)
Daily Storj S3 ≤ 24 h
Off-site cold copy Weekly Storj S3 ≤ 7 d
Redis snapshot (RDB) Every 6 h Local + Storj ≤ 6 h

WAL archiving to S3-compatible storage is still pending — that's the next item. With it, RPO drops to seconds. Without it, hourly logical dumps are the floor.

A backup you've never restored from is a hope, not a backup. We restore from yesterday's hourly dump into a scratch container monthly. The first time we tried, the restore script had bit-rotted and didn't compile.

DELETE

, in staging, against real data, and read the results" is thirty seconds of work. It is also the only thing that would have caught this.We've run the new contract for two weeks now. No range-DELETE incidents. The new auth_source = 'test'

filter is boring and explicit and impossible to fat-finger. Boring is the goal.

This postmortem is from production work at ** HoneyChat** — a Telegram-native AI companion. Canonical version:

HoneyChat Engineering

pg_dump

documentation

── more in #ai-products 4 stories · sorted by recency
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/we-deleted-10-real-u…] indexed:0 read:6min 2026-05-28 ·