Re: [RFC PATCH 19/21] usage API: use C99 macros for {usage,usagef,die,error,warning,die}*()

Just adding another datapoint...

At my $DAYJOB, we've been collecting trace2 information but have only
used it very lightly so far.  We've been meaning to use it more.
Anyway, I have a couple (possibly not fully informed) opinions based
on that...

On Tue, Dec 28, 2021 at 8:32 AM Jeff Hostetler [off-list ref] wrote:

On 12/27/21 6:01 PM, Ævar Arnfjörð Bjarmason wrote:

quoted

On Mon, Dec 27 2021, Jeff Hostetler wrote:

quoted

On 11/15/21 5:18 PM, Ævar Arnfjörð Bjarmason wrote:
[...]

quoted

It might be a good change to remove the "fmt" key from the "error"
events as a follow-up change. As these few examples from running the
test suite show it's sometimes redundant (same as the "msg"), rather
useless (just a "%s"), or something we could now mostly aggregate by
file/line instead of the normalized printf format:
        1 file":"builtin/gc.c","line":1391,"msg":"'bogus' is not a
valid task","fmt":"'%s' is not a valid task"}
        1 file":"builtin/for-each-ref.c","line":89,"msg":"format: %(then) atom used more than once","fmt":"%s"}
        1 file":"builtin/fast-import.c","line":411,"msg":"Garbage after mark: N :202 :302x","fmt":"Garbage after mark: %s"}
"Mostly" here assumes that it would be OK if the aggregation changed
between git versions, which may be what all users of trace2 want. The
change that introduced the "fmt" code was ee4512ed481 (trace2: create
new combined trace facility, 2019-02-22), and the documentation change
was e544221d97a (trace2: Documentation/technical/api-trace2.txt,
2019-02-22).
Both are rather vague on what problem "fmt" solved exactly, aside
from
the obvious one of it being impossible to do meaningful aggregations
due to the "file" and "line" being the same everywhere, which isn't
the case now.
In any case, let's leave "fmt" be for now, the above summary was
given
in case it's interesting to remove it in the future, e.g. to save
space in trace2 payloads.

I added the "fmt" field so that we could do aggregations
of error messages across multiple users without regard
to what branch or filename or percentage or whatever was
formatted into the actual "msg" written to stderr.

The actual file:line wasn't useful (primarily because it
was probably something in usage.c), but even if we fix that
it might not be useful if we have users running 10 different
versions of Git (because some people don't upgrade immediately).

So I'd rather not kill it right now.

Thanks. I'm not trying to kill it, but just poking at what it was for
exactly.

Depending on the answer to that perhaps we didn't need it anymore, but
the explanation you provide (mostly) makes sense.

The "mostly" being because I'm assuming that you only need to deal with
LC_ALL=C users?

I.e. the documented promise that you can group things by "fmt" doesn't
hold if you're processing even streams from users who are using a
translated git, because we'll get the translated format string, not the
original.

I just did a query on the data we've collected over the last
few weeks and there are only English error messages in the
database, so yes LC_ALL=C seems to be the norm.

quoted

For that we'd need to change the API from/to to:

     - error(_("some format %s"), ...)
     + error(N_("some format %s"), ...)

So no, I don't think it is worth the complexity to change
this.  Besides, wouldn't you need to more machinery under
the hood -- to emit the untranslated string to trace2 and
the translated string to stderr?  (As in, move the translation
down a layer??)

My "fmt" field is not worth that effort.

And besides, my goal was only to get the "top 10 or 20 errors"
across a large set of users.  I guess I'm thinking of it as a
sample rather than an exhaustive list, so it is OK if we don't
capture the translated strings.

This makes sense to me.

Something else to consider is the GDPR.  The "fmt" string is
generic (e.g. "path '%s' exists on disk, but not in the index")
but doesn't leak an PII or otherwise sensitive data.  Whereas
the corresponding "msg" field does include the pathname in this
example.  So if someone is post-processing the data and aggregating,
they may want to relay only the "fmt" field and not the "msg" field
to their database.

(Granted, there are lots of PII and GDPR problematic fields in
the data stream that a post-processor would need to be aware of,
but all of that is outside of the scope of the Trace2 logging.
I only mention it here because the "fmt" field may be useful for
reasons not previously discussed.)

quoted

So being able to say "just group on file/line" would be simpler.

And also "mostly" because the "fmt" case also won't handle these and
other duplicate formats (but maybe you haven't run into them in
practice):

     $ git grep -E '\b(usage|die|error|warning)(_errno)?\("%s\"' -- '*.[ch]' | wc -l
     90

So I was somewhat hoping for future work that you'd be OK with the new
file/line grouping.

Because keeping "fmt" would eventually need some massive coccinelle
search/replacement for "_(...)" -> "N_(...)" per the above, even then
consumers of the stream would get duplicate grouping for the likes of
"%s".

Do you think if as a follow-up we had "__func__"[1] along with
"file/line" that the "file/__func__" combination would be good enough?
The advantage of that would be that we could punt that "fmt"
change/complexity and document:

     If you'd like to group errors the "file/line" pair will be unique
     enough within a given git version to do so (sans a few codepaths that
     relay errors from elsewhere).

I don't actually like the file/line pair technique so much.  We have a
little influence but no control over deployed git versions.  We've
collected data about git versions in use, and we've got every single
git version going back several years in use (plus, some people took
the copy of Git I built with ort and other stuff included early but
they haven't all upgraded as I've updated it leaving us with an even
greater sprawl of file/line numbers).  That means file/line pairs
would split messages much more than the number of languages would.
(We do have developers in over a dozen different countries, though we
have a strong majority in English-speaking countries.)

quoted

     If you'd like a semi-stable grouping across similar git versions the
     "file/func" pair should be Good Enough for most purposes. Some functions
     might emit multiple errors, but you'd probably want to group them as
     similar enough anyway.

Why would we want to group different errors?  Isn't the point to
figure out which error is being triggered the most (or which errors)?
This sounds like it'd leave us with more investigation work to do.

quoted

But I realize that those things don't give you exactly the same things
that "fmt" does, but maybe they're good enough (or even better?), or
not.

I think "fmt" is strictly better, personally.