在 postgres 中删除和插入的竞争条件

Question

I am working on a node project where I have imported pg library for database operations. I have a Kafka queue from where I fetch events and store them in the database. I am fetching orders from kafka and every time an order is updated a new event is generated, I need to delete the old order details and replace them with new ones.

Below is the code

async function saveOrders(orders: Array<TransactionOnOrders>) {
  const client = await pool.connect()
 
  try {
    await client.query('BEGIN')
    if (orders.length) {
      const deleted = await deleteOrders(client, orders[0].orderId)
      logger.debug(`deleted rowCount ${deleted.rowCount} ${orders[0].orderId}`)
    }
    const queries = orders.map(ord => saveTransactionOnOrders(client, ord))
    await Promise.all(queries)
    await client.query('COMMIT')
  } catch (e) {
    await client.query('ROLLBACK')
    throw e
  } finally {
    client.release()
  }
}

The orders are getting updated very frequently and we are receiving lots of events creating a race condition leading to records not being deleted and insertion of extra records. For eg: let's say we received an event for order123 and the transaction is in process, till the time it completes another event for order123 is received so the deletion query returns 0 rows affected and the insertion query inserts another row leading to 2 rows while there should be only one record present.

I have tried to change the isolation level that didn't work well and resulted in an error

await client.query('BEGIN TRANSACTION ISOLATION LEVEL REPEATABLE READ')
await client.query('BEGIN TRANSACTION ISOLATION LEVEL SERIALIZABLE')

Is there any mistake I am doing here or is there a better to handle the above situation?

Answer 1

This may be easier if you were updating rows rather than deleting them and recreating them. In that situation you can rely on row locks preventing concurrent updates.

Answer 2

Use INSERT ... ON CONFLICT to “upsert” the incoming rows. That is atomic and free from race conditions.

Answer 3

As others have suggested, the ideal option here is to use INSERT ... ON CONFLICT to do this atomically. I can't help with that without seeing the contents of deleteOrders and saveTransactionOrders .

If that's not an option, you should use SERIALIZABLE as the isolation level. You will then get some serialisation errors, but these are safe to retry. If you used @databases ( https://www.atdatabases.org/docs/pg-guide-transactions ) you could retry just by passing retrySerializationFailures: true :

async function saveOrders(orders: Array<TransactionOnOrders>) {
  await pool.tx(async client => 
    if (orders.length) {
      const deleted = await deleteOrders(client, orders[0].orderId)
      logger.debug(`deleted rowCount ${deleted.rowCount} ${orders[0].orderId}`)
    }
    const queries = orders.map(ord => saveTransactionOnOrders(client, ord))
    await Promise.all(queries)
  }, {
    isolationLevel: IsolationLevel.SERIALIZABLE,
    retrySerializationFailures: true,
  })
}

@databases handles starting the transaction, and committing/rolling back when the async callback ends. It also retries on serialisation failures.

If you are dealing with a very high volume of events, you may encounter performance issues due to the high frequency of serialization failures and therefore retries.

You can use a "lock" in node.js to ensure that only one process at a time updates a given set of orders. https://www.atdatabases.org/docs/lock should make this fairly simple to implement. This will only lock out concurrent transactions on a single node.js process though, so you still need the transaction handling to deal with multiple node.js processes.