y1chi commented on a change in pull request #13350:
URL: https://github.com/apache/beam/pull/13350#discussion_r524486089
##########
File path: sdks/python/apache_beam/io/mongodbio.py
##########
@@ -241,6 +275,27 @@ def _get_split_keys(self, desired_chunk_size_in_mb,
start_pos, end_pos):
max={'_id': end_pos},
maxChunkSize=desired_chunk_size_in_mb)['splitKeys'])
+ def _get_buckets(self, desired_chunk_size, start_pos, end_pos):
+ if start_pos >= end_pos:
+ # single document not splittable
+ return []
+ size = self.estimate_size()
+ bucket_count = size // desired_chunk_size
Review comment:
The split function will likely be called recursively for dynamic
rebalancing, so for a range with start_pos and end_pos, it can be further split
upon backend request, so it might not reasonable to always use the total
collection size divide by desired_chunk_size to calculate the bucket count. Is
it possible to only get the buckets within the give _id range? and we can
probably use an average document size times the number of documents to
calculate the size of the range being split.
##########
File path: sdks/python/apache_beam/io/mongodbio.py
##########
@@ -241,6 +275,27 @@ def _get_split_keys(self, desired_chunk_size_in_mb,
start_pos, end_pos):
max={'_id': end_pos},
maxChunkSize=desired_chunk_size_in_mb)['splitKeys'])
+ def _get_buckets(self, desired_chunk_size, start_pos, end_pos):
+ if start_pos >= end_pos:
+ # single document not splittable
+ return []
+ size = self.estimate_size()
+ bucket_count = size // desired_chunk_size
+ if size % desired_chunk_size != 0:
+ bucket_count += 1
+ with beam.io.mongodbio.MongoClient(self.uri, **self.spec) as client:
+ buckets = list(
Review comment:
the return buckets should guarantee the _id range is start_pos and
end_pos otherwise same document could be read multiple times.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
