https://github.com/voldemort/voldemort
Tip revision: e63bc530ca12160723246ef934ea61e877f1263e authored by ARUNACHALAM THIRUPATHI on 06 June 2015, 00:07:15 UTC
Releasing Voldemort build 1.9.15
Releasing Voldemort build 1.9.15
Tip revision: e63bc53
PREUPGRADE_FOR_1_1_X_README
This directory contains utility to convert BDB JE data between different versions of Voldemort.
Need for Conversion
-------------------
Voldemort has been using "sorted duplicates" feature of BDB JE to handle
conflicting writes to the same key. At the very minimum, the conversion gets
rid of BDB sorted duplicates support and handles duplicates in the Voldemort
storage layer itself. The decision was made after months of closely working
with Oracle JE team, to understand the factors affecting performance.
Data Formats
------------
This section describes the data formats themselves.
1) Base Format (Base)
---------------------
This is the format used by Voldemort up until 1.1.x, relying on BDB JE for
duplicate handling
Disadvantages:
-- The manner in which BDB JE handles duplicates is not suitable for an
application with small percent of 2-3 duplicates i.e Voldemort.
-- Data bloat issue that prevented us from migrating to any higher 4.x version
to be able to control cache eviction
-- Incompatible with how duplicates are handled in JE5.
-- May incur additional locking costs for the "duplicates" subtree
2) New duplicate format (NewDup)
--------------------------------
This format is supported from release 1.1.x, where Voldemort storage layer
handles duplicates and BDB JE version is bumped up to JE 4.1.17
Advantages:
-- Ability to move data off disk. This is very GC friendly, relying on OS page
cache for the data and using the JVM heap only for index. This is achieved
by setting "bdb.cache.evictln" server parameter to "true"
-- Ability to evict data brought into the cache during scans, minimize impact
on online traffic (Restore, Rebalance, Retention). This is achieved by
setting "bdb.minimize.scan.impact" to "true"
-- Thinner storage layer. eg: BdbStorageEngine.put() does not incur the cost
of an additional delete()
-- General speed up due to elimination of duplicates
This format is the minimum requirement to be able to upgrade to 1.1.x & higher
3) Partition Scan format (PidScan)
----------------------------------
This is a super set of 'NewDup' format, supported 1.1.x upwards. In addition to
eliminating duplicates and upgrading to JE 4.1.17, it adds a 2 byte prefix
representing the partition id to each key.
Advantages:
-- Speed up Restore and Rebalancing linearly to the number of partitions
actually fetched. (which means much shorter degraded mode performance)
This is an optional format. You can turn it off, by setting
bdb.prefix.keys.with.partitionid=false, if you don't like for some reason
Note : We have not seen the extra 2 bytes cause any overhead to online
performance
IMPORTANT: IT IS REQUIRED TO CONVERT TO EITHER 'NewDup' or 'PidScan' TO RUN
VOLDEMORT WITH BDB, STARTING RELEASE 1.1.x
4) BDB 5 Support
------------------------------------
From versions 1.6.0+, Voldemort has moved to BDB JE 5, offering superior
compaction and generally faster online operations.
If you are upgrading from a version prior to 1.1.x to 1.6.0+, then you
also need to include a step to preupgrade from BDB 4 to BDB 5 before
running the conversion utility below. (See BDB JE documentation for details)
Running the Conversion Utility
------------------------------
The tool provides the ability to convert one database from a source environment
to a destination environment. You need to run the tool for each of the databases
or voldemort store you have. You can bring one Voldemort server at a time and
perform the conversion and bring it up on the appropriate release
Note: For users running with "bdb.one.env.per.store=false", it means you will
have to run the tool with the same --src --dest options for each database
contained.
In addition to BDB environment locations, the tool needs the cluster.xml to generate
the partition prefix.
$./voldemort-convert-bdb.sh --src <Required: Path to source bdb environment>
--dest <Required: Path to place converted new environment>
--store <Required: BDB database (voldemort store) name>
--cluster-xml <Required: Path to cluster.xml>
--from-format <Required: Format to convert FROM, one of the 3
strings 'Base','NewDup','PidScan'>
--to-format <Required: Format to convert TO, one of the 3
strings 'Base','NewDup','PidScan'>
--je-log-size <Optional: Size in MB for the new JE log files,
Default:60>
--btree-nodemax <Optional: Btree fan out, Default: 512>
We recommend you run the following to move to release 1.1.x & up.
$./voldemort-convert-bdb.sh --src /path/to/src/env
--dest /path/to/dest/env
--store teststore
--cluster-xml /path/to/cluster/xml
--from-format Base
--to-format PidScan