def _clean_cdp_filestore(google_creds_path: Path) -> None:
# Connect to database
fs = GCSFileSystem(token=str(google_creds_path))
# Open the key to get the project id
with open(google_creds_path, "r") as open_resource:
creds = json.load(open_resource)
project_id = creds["project_id"]
# Remove all files in bucket
bucket = f"{project_id}.appspot.com"
log.info(f"Cleaning bucket: {bucket}")
try:
fs.rm(f"{bucket}/*")
# Handle empty bucket
except FileNotFoundError:
pass
log.info("Filestore cleaning complete")
class GCSFS(Operations):
def __init__(self, path='.', gcs=None, **fsargs):
if gcs is None:
self.gcs = GCSFileSystem(**fsargs)
else:
self.gcs = gcs
self.cache = {}
self.counter = 0
self.root = path
def getattr(self, path, fh=None):
try:
info = self.gcs.info(''.join([self.root, path]))
except FileNotFoundError:
raise FuseOSError(ENOENT)
data = {'st_uid': 1000, 'st_gid': 1000}
perm = 0o777
if info['storageClass'] == 'DIRECTORY' or 'bucket' in info['kind']:
data['st_atime'] = 0
data['st_ctime'] = 0
data['st_mtime'] = 0
data['st_mode'] = (stat.S_IFDIR | perm)
data['st_size'] = 0
data['st_blksize'] = 0
else:
data['st_atime'] = str_to_time(info['timeStorageClassUpdated'])
data['st_ctime'] = str_to_time(info['timeCreated'])
data['st_mtime'] = str_to_time(info['updated'])
data['st_mode'] = (stat.S_IFREG | perm)
data['st_size'] = info['size']
data['st_blksize'] = 5 * 2**20
data['st_nlink'] = 1
return data
def readdir(self, path, fh):
path = ''.join([self.root, path])
files = self.gcs.ls(path)
files = [f.rstrip('/').rsplit('/', 1)[1] for f in files]
return ['.', '..'] + files
def mkdir(self, path, mode):
bucket, key = core.split_path(path)
if not self.gcs.info(path):
self.gcs.dirs['bucket'].append({
'bucket': bucket,
'kind': 'storage#object',
'size': 0,
'storageClass': 'DIRECTORY',
'name': path.rstrip('/') + '/'
})
def rmdir(self, path):
info = self.gcs.info(path)
if info['storageClass':'DIRECTORY']:
self.gcs.rm(path, False)
def read(self, path, size, offset, fh):
print('read', path, size, offset, fh)
fn = ''.join([self.root, path])
f = self.cache[fn]
f.seek(offset)
out = f.read(size)
return out
def write(self, path, data, offset, fh):
print('write', path, offset, fh)
f = self.cache[fh]
f.write(data)
return len(data)
def create(self, path, flags):
print('create', path, oct(flags))
fn = ''.join([self.root, path])
self.gcs.touch(
fn) # this makes sure directory entry exists - wasteful!
# write (but ignore creation flags)
f = self.gcs.open(fn, 'wb')
self.cache[self.counter] = f
self.counter += 1
return self.counter - 1
def open(self, path, flags):
print('open', path, oct(flags))
fn = ''.join([self.root, path])
if flags % 2 == 0:
# read
f = self.gcs.open(fn, 'rb')
else:
# write (but ignore creation flags)
f = self.gcs.open(fn, 'wb')
self.cache[self.counter] = f
self.counter += 1
return self.counter - 1
def truncate(self, path, length, fh=None):
print('truncate', path, length, fh)
fn = ''.join([self.root, path])
if length != 0:
raise NotImplementedError
# maybe should be no-op since open with write sets size to zero anyway
self.gcs.touch(fn)
def unlink(self, path):
print('delete', path)
fn = ''.join([self.root, path])
try:
self.gcs.rm(fn, False)
except (IOError, FileNotFoundError):
raise FuseOSError(EIO)
def release(self, path, fh):
print('close', path, fh)
try:
f = self.cache[fh]
f.close()
self.cache.pop(fh, None) # should release any cache memory
except Exception as e:
print(e)
return 0
def chmod(self, path, mode):
raise NotImplementedError
class GCSFS(Operations):
def __init__(self, path='.', gcs=None, nfiles=10, **fsargs):
if gcs is None:
# minimum block size: still read on 5MB boundaries.
self.gcs = GCSFileSystem(block_size=30 * 2 ** 20,
cache_timeout=6000, **fsargs)
else:
self.gcs = gcs
self.cache = SmallChunkCacher(self.gcs, nfiles=nfiles)
self.write_cache = {}
self.counter = 0
self.root = path
@_tracemethod
def getattr(self, path, fh=None):
path = ''.join([self.root, path])
try:
info = self.gcs.info(path)
except FileNotFoundError:
parent = path.rsplit('/', 1)[0]
if path in self.gcs.ls(parent):
info = True
else:
raise FuseOSError(ENOENT)
data = {'st_uid': 1000, 'st_gid': 1000}
perm = 0o777
if (info is True or info['storageClass'] == 'DIRECTORY'
or 'bucket' in info['kind']):
data['st_atime'] = 0
data['st_ctime'] = 0
data['st_mtime'] = 0
data['st_mode'] = (stat.S_IFDIR | perm)
data['st_size'] = 0
data['st_blksize'] = 0
else:
data['st_atime'] = str_to_time(info['timeStorageClassUpdated'])
data['st_ctime'] = str_to_time(info['timeCreated'])
data['st_mtime'] = str_to_time(info['updated'])
data['st_mode'] = (stat.S_IFREG | perm)
data['st_size'] = info['size']
data['st_blksize'] = 5 * 2**20
data['st_nlink'] = 1
return data
@_tracemethod
def readdir(self, path, fh):
path = ''.join([self.root, path])
logger.info("List {}, {}".format(path, fh))
files = self.gcs.ls(path)
files = [os.path.basename(f.rstrip('/')) for f in files]
return ['.', '..'] + files
@_tracemethod
def mkdir(self, path, mode):
path = ''.join([self.root, path])
logger.info("Mkdir {}".format(path))
parent, name = path.rsplit('/', 1)
prefixes = self.gcs._listing_cache[parent + '/'][1]['prefixes']
if name not in prefixes:
prefixes.append(name)
return 0
@_tracemethod
def rmdir(self, path):
info = self.gcs.info(path)
if info['storageClass': 'DIRECTORY']:
self.gcs.rm(path, False)
@_tracemethod
def read(self, path, size, offset, fh):
fn = ''.join([self.root, path])
logger.info('read #{} ({}) offset: {}, size: {}'.format(
fh, fn, offset, size))
out = self.cache.read(fn, offset, size)
return out
@_tracemethod
def write(self, path, data, offset, fh):
fn = ''.join([self.root, path])
logger.info('write #{} ({}) offset'.format(fh, fn, offset))
f = self.write_cache[fh]
f.write(data)
return len(data)
@_tracemethod
def create(self, path, flags):
fn = ''.join([self.root, path])
logger.info('create {} {}'.format(fn, oct(flags)))
self.gcs.touch(fn) # this makes sure directory entry exists - wasteful!
# write (but ignore creation flags)
f = self.gcs.open(fn, 'wb')
self.write_cache[self.counter] = f
logger.info('-> fh #{}'.format(self.counter))
self.counter += 1
return self.counter - 1
@_tracemethod
def open(self, path, flags):
fn = ''.join([self.root, path])
logger.info('open {} {}'.format(fn, oct(flags)))
if flags % 2 == 0:
# read
self.cache.open(fn)
else:
# write (but ignore creation flags)
self.gcs.open(fn, 'wb')
self.write_cache[self.counter] = f
logger.info('-> fh #{}'.format(self.counter))
self.counter += 1
return self.counter - 1
@_tracemethod
def truncate(self, path, length, fh=None):
fn = ''.join([self.root, path])
logger.info('truncate #{} ({}) to {}'.format(fh, fn, length))
if length != 0:
raise NotImplementedError
# maybe should be no-op since open with write sets size to zero anyway
self.gcs.touch(fn)
@_tracemethod
def unlink(self, path):
fn = ''.join([self.root, path])
logger.info('delete', fn)
try:
self.gcs.rm(fn, False)
except (IOError, FileNotFoundError):
raise FuseOSError(EIO)
@_tracemethod
def release(self, path, fh):
fn = ''.join([self.root, path])
logger.info('close #{} ({})'.format(fh, fn))
try:
if fh in self.write_cache:
# write mode
f = self.write_cache[fh]
f.close()
self.write_cache.pop(fh, None)
except Exception as e:
logger.exception("exception on release:" + str(e))
return 0
@_tracemethod
def chmod(self, path, mode):
raise NotImplementedError
class DeltaReaderAppendTest(TestCase):
@classmethod
def setUpClass(self):
self.path = f"tests/{str(uuid.uuid4())}/table1"
self.spark = (
pyspark.sql.SparkSession.builder.appName("deltalake").config(
"spark.jars.packages",
"io.delta:delta-core_2.12:0.7.0").config(
"spark.sql.extensions",
"io.delta.sql.DeltaSparkSessionExtension").config(
"spark.sql.catalog.spark_catalog",
"org.apache.spark.sql.delta.catalog.DeltaCatalog",
).getOrCreate())
df = (self.spark.range(0,
1000).withColumn("number", rand()).withColumn(
"number2",
when(col("id") < 500, 0).otherwise(1)))
for i in range(12):
df.write.partitionBy("number2").format("delta").mode(
"append").save(self.path)
self.fs = GCSFileSystem(project=GCP_PROJECT_ID)
self.fs.upload(self.path, f"{GCP_BUCKET}/{self.path}", recursive=True)
self.table = DeltaTable(f"{GCP_BUCKET}/{self.path}",
file_system=self.fs)
@classmethod
def tearDownClass(self):
# remove folder when we are done with the test
self.fs.rm(f"{GCP_BUCKET}/{self.path}", recursive=True)
shutil.rmtree(self.path)
def test_paths(self):
assert self.table.path == f"{GCP_BUCKET}/{self.path}"
assert self.table.log_path == f"{GCP_BUCKET}/{self.path}/_delta_log"
def test_versions(self):
assert self.table.checkpoint == 10
assert self.table.version == 11
def test_data(self):
# read the parquet files using pandas
df_pandas = self.table.to_pandas()
# read the table using spark
df_spark = self.spark.read.format("delta").load(self.path).toPandas()
# compare dataframes. The index may not be the same order, so we ignore it
assert_frame_equal(
df_pandas.sort_values("id").reset_index(drop=True),
df_spark.sort_values("id").reset_index(drop=True),
)
def test_version(self):
# read the parquet files using pandas
df_pandas = self.table.as_version(5, inplace=False).to_pandas()
# read the table using spark
df_spark = (self.spark.read.format("delta").option(
"versionAsOf", 5).load(self.path).toPandas())
# compare dataframes. The index may not be the same order, so we ignore it
assert_frame_equal(
df_pandas.sort_values("id").reset_index(drop=True),
df_spark.sort_values("id").reset_index(drop=True),
)
def test_partitioning(self):
# Partition pruning should half number of rows
assert self.table.to_table(
filter=ds.field("number2") == 0).num_rows == 6000
def test_predicate_pushdown(self):
# number is random 0-1, so we should have fewer than 12000 rows no matter what
assert self.table.to_table(
filter=ds.field("number") < 0.5).num_rows < 12000
def test_column_pruning(self):
t = self.table.to_table(columns=["number", "number2"])
assert t.column_names == ["number", "number2"]
