ottertune

The thing this most makes me think of is Ottertune, that the Carnegie Mellon DB group came up with several years ago for MySQL, https://github.com/cmu-db/ottertune, and eventually commercialised https://ottertune.com/ for MySQL and Postgres. I keep thinking about this from a JVM perspective, there's so many settings you can tweak that can have such a drastic impact on performance, depending on your workload, that it seems likely there's some value in automated tuning. The main difficult there (from my probably naive perspective) is figuring out the right feedback loop.

Does it work? Seems it's a now a commercial enterprise that has sprung from a CMU DB group project. An old version's still on Github: https://github.com/cmu-db/ottertune They also had Peloton, a "self-driving SQL database management system", that's a dead project now.

There's a black box optimizer for databases, which is designed to take the user away from having to find and tweak all the hundreds of possible tunables in the database, https://github.com/cmu-db/ottertune / https://ottertune.cs.cmu.edu/ I keep wanting the same for JVM. It's probably nice that there are so many ways to adjust the JVM to best suit a very wide range of workloads, but from a practical perspective I rarely have spare time to go diving through the guts to experiment. Let me point something at a one-box, and it can tweak parameters to its hearts content, so I can export and apply to the rest of the fleet.

> I would wager that Java's GC is better just for the simple fact that it has a whole bunch of knobs you can turn based on your workload. Tuning the GC is a challenge in and of itself. When you get so many knobs to turn, it can get hard to rationalise the overall impact of any particular one. It's part of why skilled DBAs are worth their weight in gold, for example (there was a fun paper & tool from CMU a couple of years ago, OtterTune https://github.com/cmu-db/ottertune, which tries to apply machine learning to what is almost a black-box optimisation problem)

Yeah, cloud providers do advertise DBaaS as 'simple' but there's a reason they don't advertise anything on parameter tuning or say that they replace DBAs [1]. You pay them so that they give you a self-managed database, but they definitely leave a lot of work on your plate. RL-based tuning might change that soon though: https://github.com/cmu-db/ottertune [1] https://aws.amazon.com/blogs/database/part-1-role-of-the-dba...

So does Oracle DB as well. In the mysql world, mysqltuner has been kicking around for more than a decade now that spits out suggested configuration changes. I haven't had a need to use it in several years though https://github.com/major/MySQLTuner-perl There was an interesting project around applying blackbox optimisation to MySQL, called OtterTune on which a paper was published a couple of years ago, https://blog.acolyer.org/2017/08/11/automatic-database-manag.... They got some good results out of it. Haven't attempted to use it myself: https://github.com/cmu-db/ottertune There are just so many dials and switches that can be changed that can modify the performance of the database, it really has to be moving far beyond anything any one individual could possibly grok.

There's a lot of interesting work being done by Pavlo et al. at CMU. Peloton [1] is a super bleeding-edge experimental custom-built DBMS, but OtterTune [2] apparently completely works for MySQL. [1] https://pelotondb.io/ [2] https://github.com/cmu-db/ottertune

CMU is doing a bunch of research on exactly what you mentioned. [1] is a general SQL database that aims to operate "autonomously" by predicting workloads (although AFAIK its current capabilities are mostly limited to adaptively moving records between row-oriented and column-oriented storage). I think [2] is based on the same research, but is a more concrete tool for tuning MySQL and Postgres configuration. I think ultimately, it's a much less tractable problem than anyone expected. [1] http://pelotondb.io/ [2] https://github.com/cmu-db/ottertune