SQL cache: enable WAL (Write Ahead Logging) by default

[BoD]

WAL has shown performance gains when executing many independent inserts (see #5834).

[netlify]

✅ Deploy Preview for apollo-android-docs canceled.

Name	Link
🔨 Latest commit	6acf13f
🔍 Latest deploy log	https://app.netlify.com/sites/apollo-android-docs/deploys/6626815675084b00080d86a7

[martinbonnin]

No strong opinion here. Without a clear benchmark showing improvement, I'd probably err on the side of leaving things untouched. I'll leave it to you though.

[BoD]

Running this particular benchmark a few times, I can see a ~12% improvement when using WAL (same with the incubating version of the same test). Confidence is not ultra high however because I get a 8-15% variation between each test.

[eduardb]

Running this particular benchmark a few times, I can see a ~12% improvement when using WAL (same with the incubating version of the same test). Confidence is not ultra high however because I get a 8-15% variation between each test.

Sorry for the drive-by, but we are currently focused on app performance improvements, so we are quite invested in these topics. Just curious, what device did you run the benchmark on? Could there have been external factors on the device that introduced the variation?

Btw, there's also this great article from P-Y that could help: https://blog.p-y.wtf/statistically-rigorous-android-macrobenchmarks.

Also, using a rooted device should help, at least according to the documentation: https://developer.android.com/topic/performance/benchmarking/microbenchmark-overview#lock-clocks.

I hope this helps, and I am not being annoying xD

[BoD]

Hey @eduardb no worries at all, quite the contrary! Advice on this topic is very much welcome and appreciated. This is also a hot topic for us currently.

Thanks for the article link, a very good one indeed. The spreadsheet template is promising.

My methodology was a lot more naive: ran the test 6 times each (with/without WAL), discarded the lowest and highest values, and looked at the averages.
Benchmarks ran on a (non rooted) Samsung S10. My hunch is that cpu throttling is probably the cause of the variance. Airplane mode to avoid external apps waking up is probably a good advice too.

[BoD]

Well, after running the benchmark (a tweaked version of this one) 40 times, and with the help of the template from PY's article, I see an improvement of only ~1%, which is not really significant (results here).

Closing this one - users can still opt-in for WAL in their app if wanted, via the configure lambda.