I thought that was part of the problem that postgres range types solves? https://www.postgresql.org/docs/current/rangetypes.html
And no, I won't explain it in detail. I consider the spark team basically cretins, and it would ruin several of my pals businesses while transferring wealth to hedge funds, exchanges and advertising groups making solutions to this problem available for free. I'll continue to snigger up my sleeve at the fact that decades of open source development still haven't noticed things which were obvious in the 1960s.
Edit add: a great example of this kind of thing, if not this exact thing: some poor guy trying to solve a 25tb problem using spark, parquet files ... finally solving it with ... Awk. Old tools written when 1mb was a lot of memory; sed, awk, tr, cut: vastly better than new tools.
The "real useful" part was a click further in for me.
https://github.com/BTrDB/smartgridstore/blob/master/tools/c3...
Ingress endpoints which is usually the more annoying component, rather than the exact storage format.
> As someone who generally finds "couple million rows? Sqlite is fine
There's still unsolved annoyances with analog timeseries measurements, which aren't easily solved in SQL with a columns + btree index.
So combining an airplane altitude sensor at 0.1Hz with a 10Hz fuel sensor, starts giving you trouble in SQL because the question someone might ask is "when the plane was at 31,000 feet, at 0.85 mach, what was the fuel rate", while the plane might have gone from 30,999 to 31,001 feet without ever recording the intermediate point.
Indirectly speaking, you can do range-scans for each measurement if you consider time as something that is different from a generic column value in a generic SQL engine (to join rows by temporal proximity).
The ASF proposal from IoTDB[1] might make interesting reading on the topic (the frequencies part) or the Boeing talk from XLDB 2018 [2]
[1] - https://wiki.apache.org/incubator/IoTDBProposal
[2] - https://conf.slac.stanford.edu/xldb2018/sites/xldb2018.conf....