Many years ago, when Postgres 9.6 was still in making, my coworker said to me with excitement: Hettie, Postgres will now have the ability to run queries in parallel! There can be only four parallel processes, but still, isn’t in nice?!
And I remember exactly what I’ve replied: no, I do not like this feature a bit! You know why? Because executing queries in parallel would rarely solve performance problems. In fact, if four parallel workers would solve your performance problems, they were not really problems! It can do more harm than good, because it can mask some real performance problems for a while, and then they will turn just to be more severe.
What happened then – I’ve started a new chapter of my life at Braviant, and we had Postgres 9.5 then, and then for almost 3 years it was like I never had time to stop and upgrade :). But now I have a team, so we’ve finally planned the upgrade, and since we were already four versions behind, we planned upgrade to 9.6 and immediately to PG 10.
We’ve started from our Data Warehouse. First – in the staging environment, we we’ve tested, and observed the execution for some time. And then on the mirroring instance, and only then – on production.
And then it started! Seriously, out of all data refreshes this one is only one, which is important for the OLTP instance, because it sends data to one of our client interfaces. It started to behave inconsistently, Sometimes it would be just fine. Other times, instead of about 50 seconds it has been running for an hour, and probably won’t finish if we won’t kill it. Yes, it was obvious, that something did change in the execution plan after the upgrade. But what?! From the first glance the execution plan looked the same, all HASH JOINS, which you would expect, when you join tables with no restrictive conditions.
But it was still painfully slow. What was more puzzling – I could take out of the
equation JOIN to any table, and performance would be unpredictable. After several dozen of attempts to make things run decided to take a closer look at the execution plan. And you know what I saw?! Yes, parallel execution/! Four joins were parallelized, which resulted in the execution time been really horrible. After the issue was found, the only thing left was to figure out, how to turn this feature off 🙂