Problems with zarr 2.11.0

[frezaei98]

From Tuesday, this code "inferred_ts = tsinfer.infer(sample_data)" has not run suddenly. I am facing this error " TypeError: 'int' object is not a mapping". What can I do to fix this error?

[jeromekelleher]

Hi @frezaei98 👋, thanks for the bug report.

Can you tell us what version you're using (is it git or from conda or pip)? A minimal working example would be ideal, with a small sample data file.

[hyanwong]

That is strange. Nothing has changed in the code. Do you know what line number is giving the error. Could you post a simple test case (with data) that reproduces the error for you?

(snap, Jerome, sorry!)

[frezaei98]

Can you tell us what version you're using (is it git or from conda or pip)? A minimal working example would be ideal, with a small sample data file.

Hi,
Thank you very much for your respond.
I am using the conda. But I also checked it in Google colab with pip .
I tried to run the toy example in tsinfer.readthedocs.io, but it has same error.

[frezaei98]

I tried to run the toy example in tsinfer.readthedocs.io, but it has same error.

Hi,
I tried to run the toy example in tsinfer.readthedocs.io, but it has same error.

[jeromekelleher]

I can confirm this happens in a fresh venv... investigating.

[jeromekelleher]

Very strange:

(venv) jk@tempest$ python3 tsinfer-example.py 
Traceback (most recent call last):
  File "tsinfer-example.py", line 12, in <module>
    inferred_ts = tsinfer.infer(sample_data)
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 291, in infer
    ancestors_ts = match_ancestors(
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 449, in match_ancestors
    return matcher.match_ancestors()
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 1401, in match_ancestors
    ts = self.store_output()
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 1477, in store_output
    ts = self.get_ancestors_tree_sequence()
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 1448, in get_ancestors_tree_sequence
    self.convert_inference_mutations(tables)
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 1232, in convert_inference_mutations
    metadata = _update_site_metadata(site.metadata, constants.INFERENCE_FULL)
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 150, in _update_site_metadata
    return {"inference_type": inference_type, **current_metadata}
TypeError: 'int' object is not a mapping

Somehow we're setting the metadata to 0 for sites when I run the version from a pip installation, but not when working with the git version. Very strange...

[frezaei98]

I can confirm this happens in a fresh venv... investigating.

I have been working on tsinfer for three or four months and it was very strange that it was right until Monday, but on Tuesday I want to continue my work, I got an error....

[frezaei98]

Very strange:

(venv) jk@tempest$ python3 tsinfer-example.py 
Traceback (most recent call last):
  File "tsinfer-example.py", line 12, in <module>
    inferred_ts = tsinfer.infer(sample_data)
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 291, in infer
    ancestors_ts = match_ancestors(
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 449, in match_ancestors
    return matcher.match_ancestors()
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 1401, in match_ancestors
    ts = self.store_output()
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 1477, in store_output
    ts = self.get_ancestors_tree_sequence()
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 1448, in get_ancestors_tree_sequence
    self.convert_inference_mutations(tables)
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 1232, in convert_inference_mutations
    metadata = _update_site_metadata(site.metadata, constants.INFERENCE_FULL)
  File "/home/jk/tmp/venv/lib/python3.8/site-packages/tsinfer/inference.py", line 150, in _update_site_metadata
    return {"inference_type": inference_type, **current_metadata}
TypeError: 'int' object is not a mapping

Somehow we're setting the metadata to 0 for sites when I run the version from a pip installation, but not when working with the git version. Very strange...

[jeromekelleher]

I'm going to reopen this @frezaei98 - there's definitely something wrong here. I'm not sure if it's directly a bug in tsinfer or some weird interaction with an upstream package, but it's something we need to get to the bottom of.

[jeromekelleher]

I've tracked this down to a problem with Zarr @frezaei98. Something changed in the behaviour of metadata with version 2.11.

MWE:

import tsinfer

with tsinfer.SampleData(sequence_length=6) as sample_data:
    sample_data.add_site(0, [0, 1, 0, 0, 0], ["A", "T"])

for site in sample_data.sites():
    print(site)

with version 2.10 we get:

Site(id=0, position=0.0, ancestral_state='A', metadata={}, time=nan, alleles=('A', 'T'))

and version 2.11 we have

Site(id=0, position=0.0, ancestral_state='A', metadata=0, time=nan, alleles=('A', 'T'))

Note the metadata=0 rather than metadata={}.

The short term fix is to install zarr version 2.10, e.g. python3 -m pip install zarr==2.10.

I'm not sure if this is a bug in zarr or whether we were depending on some undocumented/dodgy behaviour. I guess we need to figure this out before we decide what the right approach is.

[hyanwong]

Thanks for tracking this down @jeromekelleher. Good bug hunting there.

[benjeffery]

Bisected to zarr-developers/zarr-python@f461eb7 still figuring out which bit of it! Seems related to handling of empty chunks.

[hyanwong]

Good luck @benjeffery !

[jeromekelleher]

Aha, fantastic, thanks @benjeffery! I bet it's the write_empty_chunks=False default that's hit us, and we just have to set that to True explicitly. I'll open a PR to test.

[benjeffery]

Aha, fantastic, thanks @benjeffery! I bet it's the write_empty_chunks=False default that's hit us, and we just have to set that to True explicitly. I'll open a PR to test.

Yep, I expect so. Was just trying to find the spot in tsinfer.

[benjeffery]

@jeromekelleher adding write_empty_chunks = True at formats.py:1059 hasn't worked for me. The only other place that default value is used is on read-back internally in zarr (hierarchy.py:355), where we have no control. Adding the param there does fix.

I think this is a PR to zarr, will write up.

[jeromekelleher]

Yeah, I tried to do a quick fix over in #631, which didn't work. Thanks for following up!

[benjeffery]

Reported upstream zarr-developers/zarr-python#965

[frezaei98]

.

I've tracked this down to a problem with Zarr @frezaei98. Something changed in the behaviour of metadata with version 2.11.

MWE:

import tsinfer

with tsinfer.SampleData(sequence_length=6) as sample_data:
    sample_data.add_site(0, [0, 1, 0, 0, 0], ["A", "T"])

for site in sample_data.sites():
    print(site)

with version 2.10 we get:

Site(id=0, position=0.0, ancestral_state='A', metadata={}, time=nan, alleles=('A', 'T'))

and version 2.11 we have

Site(id=0, position=0.0, ancestral_state='A', metadata=0, time=nan, alleles=('A', 'T'))

Note the metadata=0 rather than metadata={}.

The short term fix is to install zarr version 2.10, e.g. python3 -m pip install zarr==2.10.

I'm not sure if this is a bug in zarr or whether we were depending on some undocumented/dodgy behaviour. I guess we need to figure this out before we decide what the right approach is.

Many thanks. I changed the version of Zarr package and it fixed and worked.

[benjeffery]

Here's an additional zarr issue that we hit in this release:
zarr-developers/zarr-python#967