Improve validation performance by avoiding deep copy

[eslavich]

When we fixed caching of schema objects in #682, we had to continue returning unique copies of those objects from load_schema, since a certain package (ahem, jwst) was mutating some schemas and thereby corrupting the cache. That copy isn't necessary within asdf, since it doesn't mess with schemas once they're loaded, but at the time I didn't think to skip the copy step internally.

It turns out that copy adds a significant performance penalty in the validator, where schemas are loaded by tag for every tagged object in an asdf file. For files with complex trees, making deep copies about doubles the read time! So, this PR changes internal calls to load_schema to _load_schema_cached, which dodges the copy.

I tested this with the asdf file described here and on my laptop, the read time decreases from an average of 11.6 seconds to 5.7 seconds.

[codecov]

Codecov Report

Merging #784 into master will not change coverage.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master     #784   +/-   ##
=======================================
  Coverage   93.87%   93.87%           
=======================================
  Files          43       43           
  Lines        5041     5041           
=======================================
  Hits         4732     4732           
  Misses        309      309           

Impacted Files	Coverage Δ
asdf/asdf.py	92.79% <100.00%> (ø)
asdf/schema.py	94.05% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 37ab3d4...6aa1d96. Read the comment docs.

[jdavies-st]

🎉

Ed, can you file an issue over at jwst recommending not mutating the schema? I believe it was in exp_to_source, though it is possible it is doing it at a lower level in datamodels. I don't recall the details you found when investigating #682 .

The fundamental issue between the 2 packages is that asdf uses schemas to serialize/deserialize tagged objects. That's the only purpose. But jwst uses schemas to validate user data, so there's a bit of a different purpose in some sense. That said, jwst should be doing things correctly by asdf, especially if it is a performance penalty not to be doing so.

[perrygreenfield]

Can you clarify if this requires JWST to do an explicit copy of the schema, or are they calling load_schema directly thus still forcing the copy?

[eslavich]

JWST is calling load_schema directly, that happens here:

https://github.com/spacetelescope/jwst/blob/master/jwst/datamodels/model_base.py#L113

So I think this is a have-our-cake-and-eat-it-too situation. JWST shouldn't see any performance impact from the copy since it only calls load_schema once per file, whereas asdf calls the function 10,004 times when reading my complex test file.

[perrygreenfield]

Yes, it is right to make mutants feel more pain.