spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Andrew Melo <andrew.m...@gmail.com>
Subject Questions about caching
Date Tue, 11 Dec 2018 17:13:33 GMT
Greetings, Spark Aficionados-

I'm working on a project to (ab-)use PySpark to do particle physics
analysis, which involves iterating with a lot of transformations (to
apply weights and select candidate events) and reductions (to produce
histograms of relevant physics objects). We have a basic version
working, but I'm looking to exploit some of Spark's caching behavior
to speed up the interactive computation portion of the analysis,
probably by writing a thin convenience wrapper. I have a couple
questions I've been unable to find definitive answers to, which would
help me design this wrapper an efficient way:

1) When cache()-ing a dataframe where only a subset of the columns are
used, is the entire dataframe placed into the cache, or only the used
columns. E.G. does "df2" end up caching only "a", or all three
columns?

df1 = sc.read.load('test.parquet') # Has columns a, b, c
df2 = df1.cache()
df2.select('a').collect()

2) Are caches reference-based, or is there some sort of de-duplication
based on the logical/physical plans. So, for instance, does spark take
advantage of the fact that these two dataframes should have the same
content:

df1 = sc.read.load('test.parquet').cache()
df2 = sc.read.load('test.parquet').cache()

...or are df1 and df2 totally independent WRT caching behavior?

2a) If the cache is reference-based, is it sufficient to hold a
weakref to the python object to keep the cache in-scope?

3) Finally, the spark.externalBlockStore.blockManager is intriguing in
our environment where we have multiple users concurrently analyzing
mostly the same input datasets. We have enough RAM in our clusters to
cache a high percentage of the very common datasets, but only if users
could somehow share their caches (which, conveniently, are the larger
datasets), We also have very large edge SSD cache servers we use to
cache trans-oceanic I/O we could throw at this as well.

It looks, however, like that API was removed in 2.0.0 and there wasn't
a replacement. There are products like Alluxio, but they aren't
transparent, requiring the user to manually cache their dataframes by
doing save/loads to external files using "alluxio://" URIs. Is there
no way around this behavior now?

Sorry for the long email, and thanks!
Andrew

---------------------------------------------------------------------
To unsubscribe e-mail: user-unsubscribe@spark.apache.org


Mime
View raw message