Report in python native format

[gaogaotiantian]

Is there an existing proposal for this?

• I have searched the existing proposals

Is your feature request related to a problem?

Currently memray has a comprehensive CLI which is probably the recommended way to use. It also provides a context to use it inline for cases where CLI is not applicable. However, the output could only be either a file path or a socket.

In our use case, we want to track memory usage on many python runners on remote clusters - CLI is obviously not applicable. We also rely on the only communication channel we can use in our framework (which means a separate socket is not available).

In theory, we could dump to a file, read it as bytes and send it, but it seems a bit more complicated than it needs to be. Also when we read it, we probably need to dump it to a file again and utilize CLI to read it.

Is there a way to add more Python native APIs so we can save the data as a pickleable object (or json) and send it through our own channels? Also maybe a way to generate a traditional pure text report that could be used in python script directly?

Describe the solution you'd like

They easy thing would be to support at least

data = io.BytesIO()
with memray.Tracker(data):
    ...

Or

with memray.Tracker() as tracker:
    ...
data = tracker.get_data()

It would be nice to be able to generate a report like

memray.show(data)

Alternatives you considered

No response

[godlygeek]

The easy thing would be to support at least

data = io.BytesIO()
with memray.Tracker(data):

That's not really practical. We record data on each call to malloc or free, and many of those will be made in places where the GIL is not held, so we can't make calls to Python objects.

Or

with memray.Tracker() as tracker:
    ...
data = tracker.get_data()

This is technically possible, but we'd need to make a choice between buffering the entire output in memory (which might easily be gigabytes of data, depending on your tracking options and how long the tracking session lasts!) or spooling it out to disk (probably to $TMPDIR)

In theory, we could dump to a file, read it as bytes and send it, but it seems a bit more complicated than it needs to be.

It wouldn't really be any easier for Memray. We'd likely wind up needing to write to a file anyway, and read from it when you asked for the data. It wouldn't necessarily be any better than you doing it yourself explicitly, especially when it comes to interacting with things like --follow-fork mode (which writes multiple discrete and independent capture files).

It would be nice to be able to generate a report like

memray.show(data)

We have vague plans to one day expose some sort of API for programmatically generating reports. It's not an easy feature to design, from an interface point of view. Ideally users would be able to write their own new report types, but the amount of data is so massive that the performance of a reporter written in pure Python rather than an extension modules is probably unreasonable, and the cost of marshalling everything through Python objects would make all but the smallest reports untenable.