HdrHistogram of latency

kgorlo · 2017-01-06T00:36:14+00:00

Hi,I'd like to propose a feature and maybe ask for few hints on how to implement it.`lua_code_cache` should be available from `main` context so it can be...

HdrHistogram of latency

kgorlo

Hi!

I want to use access_by_lua with proxy_pass in one location and do some potentially long lasting operations there (JSON parsing for example). And I want to measure how much time I am spending there.

To store these latencies I would prefer to use https://hdrhistogram.github.io/HdrHistogram/ (there is some Lua implementation: https://github.com/hynd/lua_hdrhistogram). But I need to have one copy of this HdrHistogram structure shared between Nginx workers - is it possible to do that with Nginx Lua? I know about shared dict - but AFAIK I can only store primitives there.

Any help would be much appreciated, thanks!

Kamil

rpaprocki

Hi,

On Jan 5, 2017, at 08:36, Kamil Gorlo <kg...@gmail.com> wrote:

Hi!

I want to use access_by_lua with proxy_pass in one location and do some potentially long lasting operations there (JSON parsing for example). And I want to measure how much time I am spending there.

To store these latencies I would prefer to use https://hdrhistogram.github.io/HdrHistogram/ (there is some Lua implementation: https://github.com/hynd/lua_hdrhistogram). But I need to have one copy of this HdrHistogram structure shared between Nginx workers - is it possible to do that with Nginx Lua? I know about shared dict - but AFAIK I can only store primitives there.

Any help would be much appreciated, thanks!
Kamil

It sounds like one of the existing stapxx tools might be of use here:

https://github.com/openresty/stapxx/blob/master/README.markdown#ngx-lua-exec-time

Rather than sample and generate inline, it uses systemtap to examine the amount of time spent in Lua land and generates the histogram for you.

Also I would hope the JSON parsing (unless it's -massive- amounts of data) wouldn't take so long as to become a bottleneck ;)

And as far as storing non-primatives in a shared dict, you can always serialize your complex structure via JSON or cap n proto. The overhead might be something to be aware of, though, in a micro benchmark case.