[
https://issues.apache.org/jira/browse/HDDS-1512?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Bharat Viswanadham updated HDDS-1512:
-------------------------------------
Description:
This Jira is created to implement DoubleBuffer in OzoneManager to flush
transactions to OM DB.
h2. Flushing Transactions to RocksDB:
We propose using an implementation similar to the HDFS EditsDoubleBuffer. We
shall flush RocksDB transactions in batches, instead of current way of using
rocksdb.put() after every operation. At a given time only one batch will be
outstanding for flush while newer transactions are accumulated in memory to be
flushed later.
In DoubleBuffer it will have 2 buffers one is currentBuffer, and the other is
readyBuffer. We add entry to current buffer, and we check if another flush call
is outstanding. If not, we flush to disk Otherwise we add entries to
otherBuffer while sync is happening.
In this if sync is happening, we shall add new requests to other buffer and
when we can sync we use *RocksDB batch commit to sync to disk, instead of
rocksdb put.*
Note: If flush to disk is failed on any OM, we shall terminate the
OzoneManager, so that OM DB’s will not diverge. Flush failure should be
considered as catastrophic failure.
Scope of this Jira is to add DoubleBuffer implementation, integrating to
current OM will be done in further jira's.
was:
This Jira is created to implement DoubleBuffer in OzoneManager to flush
transactions to OM DB.
h2. Flushing Transactions to RocksDB:
We propose using an implementation similar to the HDFS EditsDoubleBuffer. We
shall flush RocksDB transactions in batches, instead of current way of using
rocksdb.put() after every operation. At a given time only one batch will be
outstanding for flush while newer transactions are accumulated in memory to be
flushed later.
In DoubleBuffer it will have 2 buffers one is currentBuffer, and the other is
readyBuffer. We add entry to current buffer, and we check if another flush call
is outstanding. If not, we flush to disk Otherwise we add entries to
otherBuffer while sync is happening.
In this if sync is happening, we shall add new requests to other buffer and
when we can sync we use *RocksDB batch commit to sync to disk, instead of
rocksdb put.*
Note: If flush to disk is failed on any OM, we shall terminate the
OzoneManager, so that OM DB’s will not diverge. Flush failure should be
considered as catastrophic failure.
> Implement DoubleBuffer in OzoneManager
> --------------------------------------
>
> Key: HDDS-1512
> URL: https://issues.apache.org/jira/browse/HDDS-1512
> Project: Hadoop Distributed Data Store
> Issue Type: Sub-task
> Reporter: Bharat Viswanadham
> Assignee: Bharat Viswanadham
> Priority: Major
>
> This Jira is created to implement DoubleBuffer in OzoneManager to flush
> transactions to OM DB.
>
> h2. Flushing Transactions to RocksDB:
> We propose using an implementation similar to the HDFS EditsDoubleBuffer. We
> shall flush RocksDB transactions in batches, instead of current way of using
> rocksdb.put() after every operation. At a given time only one batch will be
> outstanding for flush while newer transactions are accumulated in memory to
> be flushed later.
>
> In DoubleBuffer it will have 2 buffers one is currentBuffer, and the other is
> readyBuffer. We add entry to current buffer, and we check if another flush
> call is outstanding. If not, we flush to disk Otherwise we add entries to
> otherBuffer while sync is happening.
>
> In this if sync is happening, we shall add new requests to other buffer and
> when we can sync we use *RocksDB batch commit to sync to disk, instead of
> rocksdb put.*
>
> Note: If flush to disk is failed on any OM, we shall terminate the
> OzoneManager, so that OM DB’s will not diverge. Flush failure should be
> considered as catastrophic failure.
>
> Scope of this Jira is to add DoubleBuffer implementation, integrating to
> current OM will be done in further jira's.
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org
