Retry probabilistic subsampling after failure

[huddlej]

Description of proposed changes

Probabilistic subsampling can randomly fail to select any samples when the requested number of maximum samples is very small. To minimize the chances of this edge case (that also happens to affect our Travis CI tests), this PR places the probabilistic subsampling inside a loop that checks for non-zero list of selected samples and retries subsampling a fixed number of times before giving up.

With 10 maximum retries, subsampling never fails under the conditions we use in our CI test (10 groups, 5 max sequences) when tested 10,000,000 times. The resulting probability of failure is low enough that this change should resolve our flaky CI test and also prevent most users from ever encountering the same error in their own analyses.

Note that this change should not affect most people under most conditions and will mostly make Augur development more sane.

Related issue(s)

Fixes #674

Testing

I confirmed the selected maximum number of retry attempts with the following code. This code tries to recreate the effect of running augur filter 10,000,000 times to see if subsampling ever fails under the conditions used by our CI test.

import numpy as np
r = np.random.default_rng()

# Calculated sequences per group for 10 groups and
# 5 maximum sequences.
spg = 0.4882902734375
n_groups = 10
max_retry_attempts = 10

# Repeat subsampling logic many times to replicate
# Travis CI runs, for example.
for j in range(10000000):
    # Try to sample a non-zero length list for the given
    # number of groups and sequences per group. Retry at
    # most a fixed number of times.
    if not any(((r.poisson(spg, size=n_groups) > 0).sum() > 0 for i in range(max_retry_attempts))):
        print(f"Failed at {j}")
        break

[codecov]

Codecov Report

Merging #676 (68cd65a) into master (3574708) will decrease coverage by 0.00%.
The diff coverage is 7.69%.

@@            Coverage Diff             @@
##           master     #676      +/-   ##
==========================================
- Coverage   30.35%   30.34%   -0.01%     
==========================================
  Files          40       40              
  Lines        5545     5549       +4     
  Branches     1348     1349       +1     
==========================================
+ Hits         1683     1684       +1     
- Misses       3802     3805       +3     
  Partials       60       60              

Impacted Files	Coverage Δ
augur/filter.py	47.04% <7.69%> (-0.26%)	⬇️

---

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 3574708...68cd65a. Read the comment docs.

[rneher]

This looks good to me. One could probably prevent this by a more elegant sampling, but this will do the trick for most cases.