[
https://issues.apache.org/jira/browse/HADOOP-18971?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17802097#comment-17802097
]
ASF GitHub Bot commented on HADOOP-18971:
-----------------------------------------
anujmodi2021 commented on PR #6270:
URL: https://github.com/apache/hadoop/pull/6270#issuecomment-1875108638
> > I think not only footer reads but this can be expanded to other
prefetches as well.
> > Especially small files that are read fully can be cached such that
multiple streams can be catered to.
>
> i don't know how common that use is...whereas for spark/tez and workers,
reopening the same file is not unusual -they just process different parts.
>
> I think this is why prefecting doesn't do anything for orc/parquet. Note
that impala does cache the column indexes/page indexes so it doesn't need the
filesystem to secretly do it for them.
Yes, we have also had similar observations. But I feel this cross-stream
caching is a good idea for both footer reads, and small files read.
> ABFS: Enable Footer Read Optimizations with Appropriate Footer Read Buffer
> Size
> -------------------------------------------------------------------------------
>
> Key: HADOOP-18971
> URL: https://issues.apache.org/jira/browse/HADOOP-18971
> Project: Hadoop Common
> Issue Type: Sub-task
> Components: fs/azure
> Affects Versions: 3.3.6
> Reporter: Anuj Modi
> Assignee: Anuj Modi
> Priority: Major
> Labels: pull-request-available
>
> Footer Read Optimization was introduced to Hadoop azure in this Jira:
> https://issues.apache.org/jira/browse/HADOOP-17347
> and was kept disabled by default.
> This PR is to enable footer reads by default based on the results of analysis
> performed as below:
> In our scale workload analysis, it was found that workloads working with
> Parquet (or for that matter OCR etc.) have a lot of footer reads. Footer
> reads here refers to the read operations done by workload to get the metadata
> of the parquet file which is required to understand where the actual data
> resides in the parquet.
> This whole process takes place in 3 steps:
> # Workload reads the last 8 bytes of parquet file to get the offset and size
> of the metadata which is present just above these 8 bytes.
> # Using that offset, workload reads the metadata to get the exact offset and
> length of data which it wants to read.
> # Workload performs the final read operation to get the data it wants to use
> for its purpose.
> Here the first two steps are metadata reads that can be combined into a
> single footer read. When workload tries to read certain last few bytes of
> data (let's say this value is footer size), driver will intelligently read
> some extra bytes above the footer size to cater to the next read which is
> going to come.
> Q. What is the footer size of file?
> A: 16KB. Any read request trying to get the data within last 16KB of the file
> will qualify for whole footer read. This value is enough to cater to all
> types of files including parquet, OCR, etc.
> Q. What is the buffer size to read when reading the footer?
> A. Let's call this footer read buffer size. Prior to this PR footer read
> buffer size was same as read buffer size (default 4MB). It was found that for
> most of the workload required footer size was only 256KB. i.e. For almost all
> parquet files metadata for that file was found to be within last 256KBs.
> Keeping this in mind it does not make sense to read whole buffer length of
> 4MB as a part of footer read. Moreover, reading larger data than require
> incur additional costs in terms of server and network latencies. Based on
> this and extensive experimentation it was observed that footer read buffer
> size of 512KB is ideal for almost all the workloads running on parquet, OCR,
> etc.
> Following configuration was introduced to configure the footer read buffer
> size:
> {*}fs.azure.footer.read.request.size{*}: default 512 KB.
> *Quantitative Stats:* For a workload running on parquet files the number of
> read requests got reduced by 2.3M down from 20M. That means around 10%
> reduction in overall TPS.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: common-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: common-issues-h...@hadoop.apache.org
