rowCount not always equal to rows.length

[spollack]

Since upgrading to pg 4.x javascript from pg 2.11.1 native, we have seen a variety of nasty failures in production under load. one of my engineers had a theory, and it proved correct: we saw in production over the weekend a case where the results returned from pg were internally inconsistent: results.rowCount was 1, but results.rows.length was 0. Our codebase has been using these two properties interchangeably, which led to some very nasty bugs. We do not currently have a repro case for this, but seems to happen on an error path under load. This was using pg 4.5.1 javascript, plus PR #961.

I would caution everyone to avoid using the rowCount property until we get to the bottom of this.

For now we have now rolled back to pg 2.11.1 native in production.

[vitaly-t]

Property rowCount equals to the number of rows affected by the query. It is not the same as rows.length, which equals to the number of rows received.

For example, when you execute an UPDATE, you will get rowCount = number of records updated, while your rows.length would be 0.

They are NOT interchangeable.

[spollack]

Good info, thanks. We have already modified our codebase to no longer look at rowCount.

As further background, the query in question was an UPDATE ... RETURNING *. Under the native pg driver to the best of our knowledge it always returned rowCount == rows.length, but with the javascript driver under stress/failure conditions we experienced a mismatch. my suspicion is that the error handling may be somewhat different here.

[vitaly-t]

They should match in this case indeed. It would take to dissect your stress test to understand the problem. Let us know if you find anything ;)

[vitaly-t]

I have just added pg-native support into pg-promise, and would be very interested to know in which cases this error occurs.