Replace tf.logging with absl.logging

[stephanwlee]

tf.logging.* is getting deprecated. We will use absl.logging instead from now on.

[wchargin]

I don’t have a lot of context on this change, but the technical side
looks good to me.

[nfelt]

I think we'll need a "shim" dependency we can use for the internal build to add the requisite absl logging dep, like "expect_absl_logging_installed", added to all the build targets that use logging. (Mostly it will be replacing the tb/compat/tensorflow rule.) We'll want to check when we sync this internally that internal logging continues to work as expected.

Externally, we should also make absl-py a hard requirement for the pip package now that we're directly depending on it, not just using it via TF or conditionally. Let's use absl-py >= 0.4 since that's the release that includes the absl.flags.argparse_flags support as well, and we'll want to be sure we use the absl flag parsing to get the absl.logging flag control.

[stephanwlee]

@nfelt thanks a lot for catching few bugs and the great suggestion.

Btw, with absl.logging, we lost color for warn/error. Is this a big deal? If so, do you know where I can learn more about this? (can't seem to find anything about ansi colors in absl or tf_logging.py)

[nfelt]

Without looking at any actual code, my guess would be that with the old logging we were using our custom LogFormatter in util.py that did coloration, whereas with the absl.logging we're using their LogFormatter which has the same structured format but doesn't do color. I would need to look at the code to confirm though.

Done!

[stephanwlee]

@nfelt should we just format everything absl or should we change the code so we use tensorboard logger?

[nfelt]

I'm not 100% sure I understand the question, but now that you mention it, it probably would be a good idea to use specific logger instances (like tensorflow uses theirs) rather than the root logger, which makes our logging better behaved if it's embedded into other programs.

We could either just do something like this:

log = logging.getLogger('tensorboard')  # or logging.getLogger(__name__)
...
def foo():
  log.info(...)

Or we could define something like a tensorboard.util.logging module, with a get_logger() function that returns the TB logger. With that arrangement we could just put the fake absl dependency on that module, and most of our code would depend on that module instead. (We could also move this function there.)

[nfelt]

I'm not 100% sure I understand the question, but now that you mention it, it probably would be a good idea to use specific logger instances (like tensorflow uses theirs) rather than the root logger, which makes our logging better behaved if it's embedded into other programs.

We could either just do something like this:

log = logging.getLogger('tensorboard')  # or logging.getLogger(__name__)
...
def foo():
  log.info(...)

Or we could define something like a tensorboard.util.logging module, with a get_logger() function that returns the TB logger. With that arrangement we could just put the fake absl dependency on that module, and most of our code would depend on that module instead. (We could also move this function there.)

[nfelt]

Do we still want to use our own log handler here (and for the cases above)? I thought there was already the absl-provided one that gets installed at the root level. So I'd actually have thought that with this arrangement we might get double logging - does that happen?

Unless our handler is much better than the standard absl one, I'd rather just use that one to simplify things.

[stephanwlee]

About what we had discussed offline: it looks like the argument/flag has to be parsed before we can determine whether the flag is set to default. This changes when we can configure logger (I am thinking putting it inside program#configure which is a bit larger scope to do in this PR. I will create a follow up PR for that.