n3nash commented on a change in pull request #2967:
URL: https://github.com/apache/hudi/pull/2967#discussion_r636319318
##########
File path:
docs/_posts/2021-05-19-employing-right-configurations-for-hudi-cleaner.md
##########
@@ -0,0 +1,77 @@
+---
+title: "Employing correct configurations for Hudi's cleaner table service"
+excerpt: "Achieving isolation between Hudi writer and readers using
`HoodieCleaner.java`"
+author: pratyakshsharma
+category: blog
+---
+
+Apache Hudi provides snapshot isolation between writers and readers. This is
made possible by Hudi’s MVCC concurrency model. In this blog, we will explain
how to employ the right configurations to manage multiple file versions.
Furthermore, we will discuss mechanisms available to users generating Hudi
tables on how to maintain just the required number of old file versions so that
long running readers do not fail.
+
+### Reclaiming space and bounding your data lake growth
+
+Hudi provides different table management services to be able to manage your
tables on the data lake. One of these services is called the **Cleaner**. As
you write more data to your table, for every batch of updates received, Hudi
can either generate a new version of the data file with updates applied to
records (COPY_ON_WRITE) or write these delta updates to a log file, avoiding
rewriting newer version of an existing file (MERGE_ON_READ). In such
situations, depending on the frequency of your updates, the number of file
versions of log files can grow indefinitely. If your use-cases do not require
keeping an infinite history of these versions, it is imperative to have a
process that reclaims older versions of the data. This is Hudi’s cleaner
service.
+
+### Problem Statement
+
+In a data lake architecture, it is a very common scenario to have readers and
writers concurrently accessing the same table. As the Hudi cleaner service
periodically reclaims older file versions, scenarios arise where a long running
query might be accessing a file version that is deemed to be reclaimed by the
cleaner. Here, we need to employ the correct configs to ensure readers (aka
queries) don’t fail.
+
+### Deeper dive into Hudi Cleaner
+
+To deal with the mentioned scenario, lets understand the different cleaning
policies that Hudi offers and the corresponding properties that need to be
configured. Options are available to schedule cleaning asynchronously or
synchronously. Before going into more details, we would like to explain a few
underlying concepts:
+
+ - **Hudi base file**: Columnar file which consists of final data after
compaction. A base file’s name follows the following naming convention:
`<fileId>_<writeToken>_<instantTime>.parquet`. In subsequent writes of this
file, file id remains the same and commit time gets updated to show the latest
version. This also implies any particular version of a record, given its
partition path, can be uniquely located using the file id and instant time.
+ - **File slice**: A file slice consists of the base file and any log files
consisting of the delta, in case of MERGE_ON_READ table type.
+ - **Hudi File Group**: Any file group in Hudi is uniquely identified by the
partition path and the file id that the files in this group have as part of
their name. A file group consists of all the file slices in a particular
partition path. Also any partition path can have multiple file groups.
+
+### Cleaning Policies
+
+Hudi cleaner currently supports below cleaning policies:
+
+ - **KEEP_LATEST_COMMITS**: This is the default policy. This is a temporal
cleaning policy that ensures the effect of having lookback into all the changes
that happened in the last X commits. Suppose a writer ingesting data into a
Hudi dataset every 30 minutes and the longest running query can take 5 hours to
finish, then the user should retain atleast the last 10 commits. With such a
configuration, we ensure that the oldest version of a file is kept on disk for
at least 5 hours, thereby preventing the longest running query from failing at
any point in time. Incremental cleaning is also possible using this policy.
Review comment:
Nit : "Suppose a writer ingesting data into a Hudi dataset every 30
minutes" -> "Suppose a writer **is** ingesting data into a Hudi dataset every
30 minutes"
--
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]
