On 2019.07.30 14:00, Jeff Hostetler wrote: > > > On 7/29/2019 6:20 PM, Josh Steadmon wrote: > > trace2 can write files into a target directory. With heavy usage, this > > directory can fill up with files, causing difficulty for > > trace-processing systems. > > I'm routing data in my org to a daemon via a Named Pipe or UD Socket, > so I'm not seeing the thousands of files problems that you're seeing. > However, we were being overwhelmed with lots of "uninteresting" commands > and so I added some whitelisting to my post-processing daemon. For > example, I want to know about checkout and push times -- I really don't > care about rev-parse or config times or other such minor commands. > > I went one step further and allow either "(cmd_name)" or > the pair "(cmd_name, cmd_mode)". This lets me select all checkouts > and limit checkouts to branch-changing ones, for example. I drop > any events in my post-processor that does not match any of my whitelist > patterns. > > Perhaps you could run a quick histogram and see if something would > be useful to pre-filter the data. That is, if we had whitelisting > within git.exe itself, would you still have too much data and/or > would you still need the overload feature that you've proposed in > this RFC? Our problem is not so much with the volume of data as the fact that a few special users have a ton of git invocations in rapid succession, each of which creates a new trace file. If we had a more synchronous collection system like yours it would probably not be so much of a challenge. But the massive number of files created in a short timeframe revealed some inefficiencies in our collection system (which are thankfully being addressed by the owners of that code). But we probably still want some sort of overload prevention feature as long as we're using the target directory approach.
