Merge pull request #31571 from frreiss:issue-lmdb-endian

PiperOrigin-RevId: 264453617

tensorflow/contrib/data/python/kernel_tests/lmdb_dataset_op_test.py

@@ -20,6 +20,7 @@ from __future__ import print_function
 
 import os
 import shutil
+import sys
 
 from tensorflow.contrib.data.python.ops import readers
 from tensorflow.python.data.kernel_tests import test_base
@@ -40,7 +41,10 @@ class LMDBDatasetTest(test_base.DatasetTestBase):
   def setUp(self):
     super(LMDBDatasetTest, self).setUp()
     # Copy database out because we need the path to be writable to use locks.
-    path = os.path.join(prefix_path, "lmdb", "testdata", "data.mdb")
+    # The on-disk format of an LMDB database is different on big-endian
+    # machines, because LMDB is a memory-mapped database.
+    db_file = "data.mdb" if sys.byteorder == "little" else "data_bigendian.mdb"
+    path = os.path.join(prefix_path, "lmdb", "testdata", db_file)
     self.db_path = os.path.join(self.get_temp_dir(), "data.mdb")
     shutil.copy(path, self.db_path)
 

tensorflow/core/BUILD

@@ -5482,11 +5482,19 @@ filegroup(
     testonly = 1,
     srcs = [
         # A simple key-value store:
+        #   0 : 'b'
+        #   1 : 'b'
+        #    ...
+        #   9 : 'b'
+        # Which is then overwritten with:
         #   0 : 'a'
         #   1 : 'b'
         #    ...
         #   9 : 'j'
         "lib/lmdb/testdata/data.mdb",
+        # LMDB, being a memory-mapped database, uses a different file format on
+        # big-endian systems.
+        "lib/lmdb/testdata/data_bigendian.mdb",
     ],
     visibility = ["//visibility:public"],
 )

tensorflow/python/kernel_tests/reader_ops_test.py

@@ -22,6 +22,7 @@ import collections
 import gzip
 import os
 import shutil
+import sys
 import threading
 import zlib
 
@@ -754,7 +755,10 @@ class LMDBReaderTest(test.TestCase):
   def setUp(self):
     super(LMDBReaderTest, self).setUp()
     # Copy database out because we need the path to be writable to use locks.
-    path = os.path.join(prefix_path, "lmdb", "testdata", "data.mdb")
+    # The on-disk format of an LMDB file is different on big-endian machines,
+    # because LMDB is a memory-mapped database.
+    db_file = "data.mdb" if sys.byteorder == "little" else "data_bigendian.mdb"
+    path = os.path.join(prefix_path, "lmdb", "testdata", db_file)
     self.db_path = os.path.join(self.get_temp_dir(), "data.mdb")
     shutil.copy(path, self.db_path)