[
https://issues.apache.org/jira/browse/CASSANDRA-6504?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Aleksey Yeschenko updated CASSANDRA-6504:
-----------------------------------------
Description:
Continuing CASSANDRA-4775 here.
We are changing counter write path to explicitly
lock-read-modify-unlock-replicate, thus getting rid of the previously used
'local' (deltas) and 'remote' shards distinction. Unfortunately, we can't
simply start using 'remote' shards exclusively, since shard merge rules
prioritize the 'local' shards. Which is why we are introducing the third shard
type - 'global', the only shard type to be used in 2.1+.
The updated merge rules are going to look like this:
global + global = keep the shard with the highest logical clock
global + local or remote = keep the global one
local + local = sum counts (and logical clock)
local + remote = keep the local one
remote + remote = keep the shard with highest logical clock
This is required for backward compatibility with pre-2.1 counters. To make
2.0-2.1 live upgrade possible, 'global' shard merge logic will have to be back
ported to 2.0. 2.0 will not produce them, but will be able to understand the
global shards coming from the 2.1 nodes during the live upgrade. See
CASSANDRA-6505.
Other changes introduced in this issue:
1. replicate_on_write is gone. From now on we only avoid replication at RF 1.
2. REPLICATE_ON_WRITE stage is gone
3. counter mutations are running in their own COUNTER_MUTATION stage now
4. counter mutations have a separate counter_write_request_timeout setting
5. mergeAndRemoveOldShards() code is gone, for now, until/unless a better
solution is found
6. we only replicate the fresh global shard now, not the complete (potentially
quite large) counter context
7. to help with concurrency and reduce lock contention, we cache node's global
shards in a new counter cache ({cf id, partition key, cell name} -> {count,
clock}). The cache is only used by counter writes, to help with 'hot' counters
being simultaneously updated.
Improvements to be handled by separate JIRA issues:
1. Split counter context into separate cells - one shard per cell. See
CASSANDRA-6506. This goes into either 2.1 or 3.0.
Potential improvements still being debated:
1. Coalesce the mutations in COUNTER_MUTATION stage if they share the same
partition key, and apply them together, to improve the locking situation when
updating different counter cells in one partition. See CASSANDRA-6508. Will to
into 2.1 or 3.0, if deemed beneficial.
was:
Continuing CASSANDRA-4775 here.
We are changing counter write path to explicitly
lock-read-modify-unlock-replicate, thus getting rid of the previously used
'local' (deltas) and 'remote' shards distinction. Unfortunately, we can't
simply start using 'remote' shards exclusively, since shard merge rules
prioritise the 'local' shards. Which is why we are introducing the third shard
type - 'global', the only shard type to be used in 2.1+.
The updated merge rules are going to look like this:
global + global = keep the shard with the highest logical clock ({count, clock}
pair will actually be replaced with {increment count, decrement count} tuple -
see CASSANDRA-6507)
global + local or remote = keep the global one
local + local = sum counts (and logical clock)
local + remote = keep the local one
remote + remote = keep the shard with highest logical clock
This is required for backward compatibility with pre-2.1 counters. To make
2.0-2.1 live upgrade possible, 'global' shard merge logic will have to be back
ported to 2.0. 2.0 will not produce them, but will be able to understand the
global shards coming from the 2.1 nodes during the live upgrade. See
CASSANDRA-6505.
Other changes introduced in this issue:
1. replicate_on_write is gone. From now on we only avoid replication at RF 1.
2. REPLICATE_ON_WRITE stage is gone
3. counter mutations are running in their own COUNTER_MUTATION stage now
4. counter mutations have a separate counter_write_request_timeout setting
5. mergeAndRemoveOldShards() code is gone, for now, until/unless a better
solution is found
6. we only replicate the fresh global shard now, not the complete (potentially
quite large) counter context
7. to help with concurrency and reduce lock contention, we cache node's global
shards in a new counter cache ({cf id, partition key, cell name} -> {count,
clock}). The cache is only used by counter writes, to help with 'hot' counters
being simultaneously updated.
Improvements to be handled by separate JIRA issues:
1. Replace {count, clock} with {increment count, decrement count} tuple. When
merging two global shards, the maximums of both will be picked. See
CASSANDRA-6507. This goes into 2.1, and makes the new implementation match
PN-Counters from the http://hal.inria.fr/docs/00/55/55/88/PDF/techreport.pdf
white paper.
2. Split counter context into separate cells - one shard per cell. See
CASSANDRA-6506. This goes into either 2.1 or 3.0.
Potential improvements still being debated:
1. Coalesce the mutations in COUNTER_MUTATION stage if they share the same
partition key, and apply them together, to improve the locking situation when
updating different counter cells in one partition. See CASSANDRA-6508. Will to
into 2.1 or 3.0, if deemed beneficial.
> counters++
> ----------
>
> Key: CASSANDRA-6504
> URL: https://issues.apache.org/jira/browse/CASSANDRA-6504
> Project: Cassandra
> Issue Type: Improvement
> Reporter: Aleksey Yeschenko
> Assignee: Aleksey Yeschenko
> Fix For: 2.1
>
>
> Continuing CASSANDRA-4775 here.
> We are changing counter write path to explicitly
> lock-read-modify-unlock-replicate, thus getting rid of the previously used
> 'local' (deltas) and 'remote' shards distinction. Unfortunately, we can't
> simply start using 'remote' shards exclusively, since shard merge rules
> prioritize the 'local' shards. Which is why we are introducing the third
> shard type - 'global', the only shard type to be used in 2.1+.
> The updated merge rules are going to look like this:
> global + global = keep the shard with the highest logical clock
> global + local or remote = keep the global one
> local + local = sum counts (and logical clock)
> local + remote = keep the local one
> remote + remote = keep the shard with highest logical clock
> This is required for backward compatibility with pre-2.1 counters. To make
> 2.0-2.1 live upgrade possible, 'global' shard merge logic will have to be
> back ported to 2.0. 2.0 will not produce them, but will be able to understand
> the global shards coming from the 2.1 nodes during the live upgrade. See
> CASSANDRA-6505.
> Other changes introduced in this issue:
> 1. replicate_on_write is gone. From now on we only avoid replication at RF 1.
> 2. REPLICATE_ON_WRITE stage is gone
> 3. counter mutations are running in their own COUNTER_MUTATION stage now
> 4. counter mutations have a separate counter_write_request_timeout setting
> 5. mergeAndRemoveOldShards() code is gone, for now, until/unless a better
> solution is found
> 6. we only replicate the fresh global shard now, not the complete
> (potentially quite large) counter context
> 7. to help with concurrency and reduce lock contention, we cache node's
> global shards in a new counter cache ({cf id, partition key, cell name} ->
> {count, clock}). The cache is only used by counter writes, to help with 'hot'
> counters being simultaneously updated.
> Improvements to be handled by separate JIRA issues:
> 1. Split counter context into separate cells - one shard per cell. See
> CASSANDRA-6506. This goes into either 2.1 or 3.0.
> Potential improvements still being debated:
> 1. Coalesce the mutations in COUNTER_MUTATION stage if they share the same
> partition key, and apply them together, to improve the locking situation when
> updating different counter cells in one partition. See CASSANDRA-6508. Will
> to into 2.1 or 3.0, if deemed beneficial.
--
This message was sent by Atlassian JIRA
(v6.1.4#6159)
