On loading the same XTC offset in parallel

.github/actions/setup-deps/action.yaml

@@ -20,6 +20,8 @@ inputs:
     default: 'codecov'
   cython:
     default: 'cython'
+  fasteners:
+    default: 'fasteners'
   griddataformats:
     default: 'griddataformats'
   gsd:
@@ -90,6 +92,7 @@ runs:
           ${{ inputs.biopython }}
           ${{ inputs.codecov }}
           ${{ inputs.cython }}
+          ${{ inputs.fasteners }}
           ${{ inputs.griddataformats }}
           ${{ inputs.gsd }}
           ${{ inputs.hypothesis }}

.gitignore

@@ -50,3 +50,4 @@ benchmarks/results
 *~
 .idea
 .vscode
+*.lock

.travis.yml

@@ -25,7 +25,7 @@ matrix:
         - conda activate base
         - conda update --yes conda
         - conda install --yes pip pytest==6.1.2 mmtf-python biopython networkx cython matplotlib-base "scipy>=1.6" griddataformats hypothesis gsd codecov -c conda-forge -c biobuilds
-        - pip install pytest-xdist packaging threadpoolctl
+        - pip install pytest-xdist packaging threadpoolctl fasteners
         - conda info
       install:
         - (cd package/ && python setup.py develop) && (cd testsuite/ && python setup.py install)
@@ -45,7 +45,7 @@ matrix:
       if: type = cron
       before_install:
         - python -m pip install cython "numpy>=1.19.2" scipy
-        - python -m pip install --no-build-isolation hypothesis matplotlib packaging pytest pytest-cov pytest-xdist tqdm threadpoolctl
+        - python -m pip install --no-build-isolation hypothesis matplotlib packaging pytest pytest-cov pytest-xdist tqdm threadpoolctl fasteners
       install:
         - cd package
         - python setup.py install

azure-pipelines.yml

@@ -95,6 +95,7 @@ jobs:
       h5py
       tqdm
       threadpoolctl
+      fasteners
     displayName: 'Install dependencies'
   # for wheel install testing, we pin to an
   # older NumPy, the oldest version we support and that

maintainer/conda/environment.yml

@@ -6,6 +6,7 @@ dependencies:
   - chemfiles>=0.10
   - codecov
   - cython
+  - fasteners
   - griddataformats
   - gsd
   - h5py

package/CHANGELOG

@@ -16,11 +16,15 @@ The rules for this file:
 ??/??/?? IAlibay, BFedder, inomag, Agorfa, aya9aladdin, shudipto-amin, cbouy,
          HenokB, umak1106, tamandeeps, Mrqeoqqt, megosato, AnirG, rishu235,
          mtiberti, manishsaini6421, Sukeerti1, robotjellyzone, markvrma, alescoulie,
-         mjtadema, PicoCentauri, Atharva7K, aditi2906, orbeckst
+         mjtadema, PicoCentauri, Atharva7K, aditi2906, orbeckst,  yuxuanzhuang
 
  * 2.2.0
 
 Fixes
+  * Fix the competition when generating offsets from multiple processes by
+    using a interprocess file lock. (Issue #1988, PR #3375)
+  * Fix the error when trying to load a corrupted offset file (Issue #3230
+    PR #3375)  
   * Iterating a SingleFrameReaderBase trajectory now rewinds the
     reader, resetting any modified trajectory attributes to those
     read from file. Fixes behavior that was previously undefined (Issue #3423)
@@ -71,6 +75,7 @@ Enhancements
     keyword (Issue #3605)
 
 Changes
+  * `fasteners` package is now a core dependency (Issues #3230, #1988, PR #3375)
   * The minimm Python, NumPy, and SciPy versions (following NEP29) are now 3.8, 
     1.19, and 1.5.0; GSD now also has a minimum version of 1.9.3 and networkx
     has minimum version 2.0 (Issue #3665)

package/MDAnalysis/coordinates/XDR.py

@@ -38,13 +38,15 @@
 import numpy as np
 from os.path import getctime, getsize, isfile, split, join
 import warnings
+import fasteners
 
 from . import base
 from ..lib.mdamath import triclinic_box
 
 
 def offsets_filename(filename, ending='npz'):
-    """Return offset filename for XDR files. For this the filename is appended
+    """Return offset or its lock filename for XDR files.
+    For this the filename is appended
     with `_offsets.{ending}`.
 
     Parameters
@@ -82,7 +84,9 @@ def read_numpy_offsets(filename):
     """
     try:
         return {k: v for k, v in np.load(filename).items()}
-    except IOError:
+
+    #  `ValueError` is encountered when the offset file is corrupted.
+    except (ValueError, IOError):
         warnings.warn("Failed to load offsets file {}\n".format(filename))
         return False
 
@@ -113,6 +117,8 @@ class XDRBaseReader(base.ReaderBase):
 
     .. versionchanged:: 1.0.0
        XDR offsets read from trajectory if offsets file read-in fails
+    .. versionchanged:: 2.0.0
+       Add a InterProcessLock when generating offsets
 
     """
     def __init__(self, filename, convert_units=True, sub=None,
@@ -179,53 +185,68 @@ def close(self):
 
     def _load_offsets(self):
         """load frame offsets from file, reread them from the trajectory if that
-        fails"""
+        fails. To prevent the competition of generating the same offset file
+        from multiple processes, an `InterProcessLock` is used."""
         fname = offsets_filename(self.filename)
+        lock_name = offsets_filename(self.filename,
+                                     ending='lock')
 
-        if not isfile(fname):
-            self._read_offsets(store=True)
-            return
-
-        # if offsets file read correctly, data will be a dictionary of offsets
-        # if not, data will be False
-        # if False, offsets should be read from the trajectory
-        # this warning can be avoided by loading Universe like:
-        # u = mda.Universe(data.TPR, data.XTC, refresh_offsets=True)
-        # refer to Issue #1893
-        data = read_numpy_offsets(fname)
-        if not data:
-            warnings.warn("Reading offsets from {} failed, "
-                          "reading offsets from trajectory instead\n"
-                          "Consider setting 'refresh_offsets=True' "
-                          "when loading your Universe".format(fname))
+        #  check if the location of the lock is writable.
+        try:
+            with fasteners.InterProcessLock(lock_name) as filelock:
+                pass
+        except PermissionError:
+            warnings.warn(f"Cannot write lock/offset file in same location as "
+                           "{self.filename}. Using slow offset calculation.")
             self._read_offsets(store=True)
             return
 
-        ctime_ok = size_ok = n_atoms_ok = False
+        with fasteners.InterProcessLock(lock_name) as filelock:
+            if not isfile(fname):
+                self._read_offsets(store=True)
+                return
+
+            # if offsets file read correctly, data will be a dictionary of offsets
+            # if not, data will be False
+            # if False, offsets should be read from the trajectory
+            # this warning can be avoided by loading Universe like:
+            # u = mda.Universe(data.TPR, data.XTC, refresh_offsets=True)
+            # refer to Issue #1893
+            data = read_numpy_offsets(fname)
+            if not data:
+                warnings.warn("Reading offsets from {} failed, "
+                              "reading offsets from trajectory instead.\n"
+                              "Consider setting 'refresh_offsets=True' "
+                              "when loading your Universe.".format(fname))
+                self._read_offsets(store=True)
+                return
+
+            ctime_ok = size_ok = n_atoms_ok = False
 
-        try:
-            ctime_ok = getctime(self.filename) == data['ctime']
-            size_ok = getsize(self.filename) == data['size']
-            n_atoms_ok = self._xdr.n_atoms == data['n_atoms']
-        except KeyError:
-            # we tripped over some old offset formated file
-            pass
-
-        if not (ctime_ok and size_ok and n_atoms_ok):
-            warnings.warn("Reload offsets from trajectory\n "
-                          "ctime or size or n_atoms did not match")
-            self._read_offsets(store=True)
-        else:
-            self._xdr.set_offsets(data['offsets'])
+            try:
+                ctime_ok = getctime(self.filename) == data['ctime']
+                size_ok = getsize(self.filename) == data['size']
+                n_atoms_ok = self._xdr.n_atoms == data['n_atoms']
+            except KeyError:
+                # we tripped over some old offset formated file
+                pass
+
+            if not (ctime_ok and size_ok and n_atoms_ok):
+                warnings.warn("Reload offsets from trajectory\n "
+                              "ctime or size or n_atoms did not match")
+                self._read_offsets(store=True)
+            else:
+                self._xdr.set_offsets(data['offsets'])
 
     def _read_offsets(self, store=False):
         """read frame offsets from trajectory"""
+        fname = offsets_filename(self.filename)
         offsets = self._xdr.offsets
         if store:
             ctime = getctime(self.filename)
             size = getsize(self.filename)
             try:
-                np.savez(offsets_filename(self.filename),
+                np.savez(fname,
                          offsets=offsets, size=size, ctime=ctime,
                          n_atoms=self._xdr.n_atoms)
             except Exception as e:

package/requirements.txt

@@ -1,6 +1,7 @@
 biopython
 codecov
 cython
+fasteners
 griddataformats
 gsd
 hypothesis

package/setup.py

@@ -601,6 +601,7 @@ def long_description(readme):
           'tqdm>=4.43.0',
           'threadpoolctl',
           'packaging',
+          'fasteners',
           'gsd>=1.9.3',
     ]
 

testsuite/MDAnalysisTests/coordinates/test_xdr.py

@@ -26,6 +26,7 @@
 import errno
 import numpy as np
 import os
+from os.path import split
 import shutil
 import subprocess
 
@@ -741,15 +742,24 @@ def test_offsets(self, trajectory, traj):
     def test_reload_offsets(self, traj):
         self._reader(traj, refresh_offsets=True)
 
-    def test_nonexistant_offsets_file(self, traj):
-        # assert that a nonexistant file returns False during read-in
+    def test_nonexistent_offsets_file(self, traj):
+        # assert that a nonexistent file returns False during read-in
         outfile_offsets = XDR.offsets_filename(traj)
         with patch.object(np, "load") as np_load_mock:
             np_load_mock.side_effect = IOError
             saved_offsets = XDR.read_numpy_offsets(outfile_offsets)
             assert_equal(saved_offsets, False)
 
-    def test_reload_offsets_if_offsets_readin_fails(self, trajectory):
+    def test_nonexistent_offsets_file(self, traj):
+        # assert that a corrupted file returns False during read-in
+        # Issue #3230
+        outfile_offsets = XDR.offsets_filename(traj)
+        with patch.object(np, "load") as np_load_mock:
+            np_load_mock.side_effect = ValueError
+            saved_offsets = XDR.read_numpy_offsets(outfile_offsets)
+            assert_equal(saved_offsets, False)
+
+    def test_reload_offsets_if_offsets_readin_io_fails(self, trajectory):
         # force the np.load call that is called in read_numpy_offsets
         # during _load_offsets to give an IOError
         # ensure that offsets are then read-in from the trajectory
@@ -761,6 +771,19 @@ def test_reload_offsets_if_offsets_readin_fails(self, trajectory):
                 self.ref_offsets,
                 err_msg="error loading frame offsets")
 
+    def test_reload_offsets_if_offsets_readin_value_fails(self, trajectory):
+        # force the np.load call that is called in read_numpy_offsets
+        # during _load_offsets to give an ValueError (Issue #3230)
+        # ensure that offsets are then read-in from the trajectory
+        with patch.object(np, "load") as np_load_mock:
+            np_load_mock.side_effect = ValueError
+            with pytest.warns(UserWarning, match="Failed to load offsets"):
+                trajectory._load_offsets()
+            assert_almost_equal(
+                trajectory._xdr.offsets,
+                self.ref_offsets,
+                err_msg="error loading frame offsets")
+
     def test_persistent_offsets_size_mismatch(self, traj):
         # check that stored offsets are not loaded when trajectory
         # size differs from stored size
@@ -850,8 +873,17 @@ def test_persistent_offsets_readonly(self, tmpdir):
 
         filename = str(tmpdir.join(os.path.basename(self.filename)))
         # try to write a offsets file
-        self._reader(filename)
+        with (pytest.warns(UserWarning, match="Couldn't save offsets") and
+              pytest.warns(UserWarning, match="Cannot write")):
+            self._reader(filename)
         assert_equal(os.path.exists(XDR.offsets_filename(filename)), False)
+        # check the lock file is not created as well.
+        assert_equal(os.path.exists(XDR.offsets_filename(filename,
+                                                    ending='.lock')), False)
+
+    def test_offset_lock_created(self):
+        assert os.path.exists(XDR.offsets_filename(self.filename,
+                                                   ending='lock'))
 
 
 class TestXTCReader_offsets(_GromacsReader_offsets):

testsuite/MDAnalysisTests/parallelism/test_multiprocessing.py

@@ -23,11 +23,14 @@
 import multiprocessing
 
 import numpy as np
+import os
+import shutil
 import pytest
 import pickle
 from numpy.testing import assert_equal
 
 import MDAnalysis as mda
+import MDAnalysis.coordinates
 from MDAnalysis.coordinates.core import get_reader_for
 from MDAnalysis.analysis.rms import RMSD
 
@@ -85,6 +88,19 @@ def u(request):
         top, trj = request.param
         return mda.Universe(top, trj)
 
+@pytest.fixture(scope="function")
+def temp_xtc(tmp_path):
+    fresh_xtc = tmp_path / "testing.xtc"
+    shutil.copy(XTC, fresh_xtc)
+    # In principle there is no possibility that the offset for
+    # fresh_xtc exists as it is a fresh copy for every test. However,
+    # the code is left for documentation.
+    try:
+        os.remove(MDAnalysis.coordinates.XDR.offsets_filename(fresh_xtc))
+    except OSError:
+        pass
+    return fresh_xtc
+
 
 # Define target functions here
 # inside test functions doesn't work
@@ -125,6 +141,23 @@ def test_universe_unpickle_in_new_process():
     assert_equal(ref, res)
 
 
+def test_creating_multiple_universe_without_offset(temp_xtc, ncopies=3):
+    #  test if they can be created without generating
+    #  the offset simultaneously.
+    #  The tested XTC file is way too short to induce a race scenario but the
+    #  test is included as documentation for the scenario that used to create
+    #  a problem (see PR #3375 and issues #3230, #1988)
+
+    args = (GRO, str(temp_xtc))
+    with multiprocessing.Pool(2) as p:
+        universes = [p.apply_async(mda.Universe, args) for i in range(ncopies)]
+        universes = [universe.get() for universe in universes]
+
+
+    assert_equal(universes[0].trajectory._xdr.offsets,
+                 universes[1].trajectory._xdr.offsets)
+
+
 @pytest.fixture(params=[
     # formatname, filename
     ('CRD', CRD, dict()),