Re: [PATCH 2/2] trace2: randomize/timestamp trace2 targets

["Ævar Arnfjörð Bjarmason" <avarab@gmail.com>]

On Fri, Mar 15 2019, Jeff Hostetler wrote:

> On 3/13/2019 7:49 PM, Ævar Arnfjörð Bjarmason wrote:
>>
>> On Thu, Mar 14 2019, Josh Steadmon wrote:
>>
>>> When the value of a trace2 environment variable contains instances of
>>> the string "%ISO8601%", expand them into the current UTC timestamp in
>>> ISO 8601 format.
>>
>> Any reason not to just support feeding the path to strbuf_addftime(), to
>> e.g. support a daily/hourly log?
>>
>>> When the value of a trace2 environment variable is an absolute path
>>> referring to an existing directory, write output to randomly-named
>>> files under the given directory. If the value is an absolute path
>>> referring to a non-existent file and ends with a dash, use the value as
>>> a prefix for randomly named files.
>>>
>>> The random filenames will consist of the value of the environment
>>> variable (after potential timestamp expansion), followed by a 6
>>> character random string such as would be produced by mkstemp(3).
>>>
>>> This makes it more convenient to collect traces for every git
>>> invocation by unconditionally setting the relevant trace2 envvar to a
>>> constant directory name.
>>
>> Hrm, api-trace2.txt already specifies that the "sid" is going to be
>> unique, couldn't we just have some mode where we use that?
>>
>> But then of course when we have nested processes will contain slashes,
>> so we'd either run into deep nesting or need to munge the slashes, in
>> which case we might bump against a file length limit (although I haven't
>> seen process trees deeper than 3-4).
>
> Using the "sid" would be a good place to start.  Just take the final
> component in the string (after the last slash or the whole sid if there
> are no slashes).  That will give you a filename with microseconds since
> epoch of the command's start time and the PID.
>
> That should be unique, should not require random strings, and not go
> deep in the filesystem.  And it will let you correlate files between
> child and parent commands, if you need to.
>
> So maybe if GIT_TR2_* is set to a directory, we append the final portion
> of the "sid" and create a file inside that directory.
>
>>
>> Just to pry about the use-case since I'm doing similar collecting, why
>> are you finding this easier to process?
>>
>> With the current O_APPEND semantics you're (unless I've missed
>> something) guaranteed to get a single process tree in nested order,
>> whereas with this they'll all end up in separate files and you'll need
>> to slurp them up, sort the whole thing and stitch it together yourself
>> without the benefit of stream-parsing it where you can cheat a bit
>> knowing that e.g. a "reflog expire" entry is always coming after the
>> corresponding "gc" that invoked it.
>>
>
> Yes, with O_APPEND, you should get a series of events as they happen
> on the system all properly interleaved.  And see concurrent activity.
> This file should let you grep to see individual processes if you want
> to.
>
> Routing each command to a different file is fine if you want, but
> that opens you up to having to manage and delete them.
>
> Whether to have 1 file (with occasional rotation) or 1 file-per-command
> depends, I guess, on how you want to process them.
>
> I'm routing the Trace2 data to a named-pipe/socket and have a daemon
> collecting and filtering, so I have a single pathname for output and
> yet get the per-file stream handling that I think Josh is looking for.

Is the collecting code something you can share & general enough that it
might be useful for others?