Github user JoshRosen commented on the pull request:
https://github.com/apache/spark/pull/12750#issuecomment-215943176
After a bit more time in a profiler, I was able to make this 2x faster than
my previous best time.
To give a rough idea of the structure of my benchmark:
I have a folder containing a bunch of gigantic JSON documents with huge,
deeply-nested schemas. Imagine the schema being the union of a huge number of
smaller schemas, so it's pretty big and sparse.
My benchmark harness reads in the raw JSON, caches it, unions it together a
bunch of times to inflate the benchmark runtime (to avoid small constant factor
noise), then coalesces it to a single partition to avoid task-launch overheads
(and so we measure single-core performance):
```scala
val lines = spark.read.text("...").rdd
lines.cache.count()
val coalescedLines = (lines ++ lines ++ lines ++ lines ++
lines).coalesce(1)
for (i <- 1 to 10) {
val startTime = System.currentTimeMillis
spark.read.json(coalescedLines)
println(s"Took ${System.currentTimeMillis - startTime}")
}
```
Before (f5da592fc63b8d3bc09d49c196d6c5d98cd2a013), this took around 88
seconds.
After (6bf5ee634ee7e509a88e64236f18e3c6e7a07aa2), this takes about **14.5
seconds**.
Coupled with the changes in #12741, a change that led to a massive speedup
in the `reduce` / `treeAggregate` step, this should be a huge speedup vs 1.6.x.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
