Todd Lipcon

(this is on a single r5.2xlarge VM, no big expensive cluster)

Does anyone have experience LTO-optimizing tcmalloc/jemalloc into a binary? Did you get noticeable hot path (threadcache-hit) improvements from inlining?

I should add: I do think that a timeseries-specific language (not _only_ standard SQL) is nice, I just think "looks just like SQL but has different results" is not it.

My take-away from InfluxQL: if you are going to make a SQL-like language but with different semantics, change the keywords too. Using the same syntax with different semantics confuses things. (eg GROUP BY <tag> should be "FOR EACH VALUE OF <tag>" or something)

The more I look at InfluxQL, the more I understand why it's being canned in favor of Flux. The more I look at Flux, the less I understand why there's a new language being invented... interesting product strategy choices at @InfluxDB .

Considering going on Shark Tank with this idea: sour cream and onion envelopes.

Also curious if you have done any versions for 8x16 bit input

Should note that something like half the cpu is going to string splitting. Think I could double this performance with a little microoptimizing work.

Easily ingesting 3.5M metrics per second from 4 threads into my personal workstation. Simple range queries with avg less than 1ms. Details to follow in a blog, I hope.

Been playing a bit with @ApacheKudu for time series using the benchmarks from @InfluxDB and @TimescaleDB. Despite being a general purpose store, Kudu turns out very competitive so far!

Sad discovery about my new AMD processor: significantly less complete PMU features vs Intel (no LBR or 'perf mem' I guess)

building new computer. convince me in 180 characters or less: latest fedora vs latest ubuntu vs ubuntu LTS? I've run Ubuntu LTS for years, Debian testing before that. Will "bleeding edge" distro make me waste lots of time screwing with things?

(the expectation is true 99% of the time and the function is small, and likely has strong knock-on optimization potential, so really want to inline it). Seems like dynamic_cast<> is full of overhead

Anyone have a C++ trick to explicitly code speculative devirtualization, preferably that works in Clang and GCC without LTO? eg I want the following to exist: builtin_expect_type(base_ptr, SomeImplClass)->VirtualFunc()