Yes, I stand corrected, dynamic writes is now much more than the primitive window-based naming we used to have.
It would be interesting to visualize how much of this codepath is metatada vs. the actual data. In the case of file writing, it seems one could (maybe?) avoid requiring a stable input, as shards are accepted as a whole (unlike, say, sinks where a deterministic uid is needed for deduplication on retry). On Wed, Aug 22, 2018 at 4:55 PM Reuven Lax <re...@google.com> wrote: > > Robert - much of the complexity isn't due to streaming, but rather because > WriteFiles supports "dynamic" output (where the user can choose a destination > file based on the input record). In practice if a pipeline is not using > dynamic destinations the full graph is still generated, but much of that > graph is never used (empty PCollections). > > On Wed, Aug 22, 2018 at 3:12 AM Robert Bradshaw <rober...@google.com> wrote: >> >> I agree that this is concerning. Some of the complexity may have also >> been introduced to accommodate writing files in Streaming mode, but it >> seems we should be able to execute this as a single Map operation. >> >> Have you profiled to see which stages and/or operations are taking up the >> time? >> On Wed, Aug 22, 2018 at 11:29 AM Tim Robertson >> <timrobertson...@gmail.com> wrote: >> > >> > Hi folks, >> > >> > I've recently been involved in projects rewriting Avro files and have >> > discovered a concerning performance trait in Beam. >> > >> > I have observed Beam between 6-20x slower than native Spark or MapReduce >> > code for a simple pipeline of read Avro, modify, write Avro. >> > >> > - Rewriting 200TB of Avro files (big cluster): 14 hrs using Beam/Spark, >> > 40 minutes with a map-only MR job >> > - Rewriting 1.5TB Avro file (small cluster): 2 hrs using Beam/Spark, 18 >> > minutes using vanilla Spark code. Test code available [1] >> > >> > These tests were running Beam 2.6.0 on Cloudera 5.12.x clusters (Spark / >> > YARN) on reference Dell / Cloudera hardware. >> > >> > I have only just started exploring but I believe the cause is rooted in >> > the WriteFiles which is used by all our file based IO. WriteFiles is >> > reasonably complex with reshuffles, spilling to temporary files >> > (presumably to accommodate varying bundle sizes/avoid small files), a >> > union, a GBK etc. >> > >> > Before I go too far with exploration I'd appreciate thoughts on whether we >> > believe this is a concern (I do), if we should explore optimisations or >> > any insight from previous work in this area. >> > >> > Thanks, >> > Tim >> > >> > [1] https://github.com/gbif/beam-perf/tree/master/avro-to-avro
