[lirbank]

Here's a real-world example where atomic updates aren't an option - an order status transition that reads the current status from one table, validates the transition, and inserts into another:

await db().transaction(async (tx) => { await hooks?.onTxBegin?.();

  const [order] = await tx.select().from(orders)
    .where(eq(orders.id, input.id))
    .for("update");

  const [status] = await tx.select().from(orderStatuses)
    .where(eq(orderStatuses.orderId, input.id))
    .orderBy(desc(orderStatuses.createdAt))
    .limit(1);

  if (input.status === status.code)
    throw new Error("Status already set");

  await tx.insert(orderStatuses).values({ ... });

});

You need the transaction + SELECT FOR UPDATE because the validation depends on current state, and two concurrent requests could both pass the duplicate check. The hooks parameter is the barrier injection point from the article - that's how you test that the lock actually prevents the race.

[erpellan]

The standard pattern to avoid select for update (which can cause poor performance under load) is to use optimistic concurrency control.

Add a numeric version column to the table being updated, read and increment it in the application layer and use the value you saw as part of the where clause in the update statement. If you see ‘0 rows updated’ it means you were beaten in a race and should replay the operation.

[tux3]

I don't think such a broad recommendation will be good for most people, it really depends.

Optimistic updates looks great when there is no contention, and they will beat locking in a toy benchmark, but if you're not very careful they can cause insane amplification under load.

It's a similar trap as spinlocks. People keep re-discovering this great performance hack that avoids the slow locks in the standard. And some day the system has a spike that creates contention, and now you have 25 instances with 24 of them spinning like crazy, slowing to a crawl the only one that could be making progress.

It's possible to implement this pattern correctly, and it can be better in some specific situations. But a standard FOR UPDATE lock will beat the average badly implemented retry loop nine times out of ten.

[lirbank]

Good point. The barrier pattern from the article applies to both approaches - whether you're using pessimistic locks or optimistic version checks, it's good to verify that the concurrency handling actually works. Barriers let you test that your version check correctly rejects the stale update, the same way they test that your lock prevents the race.

[codys]

Seems you could use a single SQL statement for that particular formulation. Something like this, using CTEs is possible, but alternately one can reformat them as subqueries. (note: not sure how the select of orders is intended to be used, so the below doesn't use it, but it does obtain it as an expression to be used)

    WITH
     o AS (
      SELECT FROM orders
      WHERE orders.id = $1
     ),
     os AS (
      SELECT FROM orderStatuses
      WHERE orderStatuses.orderId = $1
      ORDER BY DESC orderStatuses.createdAt
      LIMIT 1
     )
     INSERT INTO orderStatuses ...
     WHERE EXISTS (SELECT 1 FROM os WHERE os.code != $2)
     RETURNING ...something including the status differ check...

Does something like this work with postgres's default behavior?

[lirbank]

Absolutely - if you can express the whole operation as a single atomic statement, that's the best outcome. No locks needed, no race to test for. The article is about what comes next: when the logic can't collapse into one query, how do you verify your concurrency handling actually works?