[
https://issues.apache.org/jira/browse/CASSANDRA-7937?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14259967#comment-14259967
]
Piotr Kołaczkowski edited comment on CASSANDRA-7937 at 12/29/14 9:58 AM:
-------------------------------------------------------------------------
I like the idea of using parallelism level as a backpressure mechanism. That
would have a nice positive effect of automatically reducing the amount of
memory used for queuing the requests.
However, I have a few concerns:
# my biggest concern is, that even limiting a single client to one write at a
time (window size = 1), might still be too fast, for some fast clients, if only
row sizes are big enough, particularly when writing big cells of data, where
big = hundreds of kB / single MBs per cell. Cassandra is extremely efficient at
ingesting data into memtables. If it was faster than we're able to flush, then
we still have a problem.
# when not using TokenAware LBP, we can have a problem that in case of
unbalanced write load, the fast (not loaded) nodes would be needlessly
penalized with lower parallelism level, because the window size would be
per-coordinator, not per partition (replica set). Am I right here?
I wonder if using a penalty delay *after* processing the write request, would
not be a better idea: in case of an unbalanced load, when one partition gets
hammered, but most others are ok, it would slow down writes for that one
partition (replica set), but would not affect latency of writes of other
partitions. I'm for applyting the delay after, because then we already know the
replica set and their load, as well as we don't need to keep data queued in
memory, for it has already been written.
The (simplified) process would look as follows:
# the write request gets accepted by the coordinator
# the write gets sent to proper replica nodes
# the replicas that acknowledge the write, also reply with their current load
information
# received load information gets averaged (median)
# when the CL is satisfied, but the load was high enough to be in the "danger
zone" the coordinator puts some artificial delay before acknowledging the write
to the client - of course small enough to not exceed the write timeout. Or if
not possible / wise for some other reasons (e.g. holding memory), just tells
the driver that "load was high, you better slow down" and the driver waits
before processing the next request.
WDYT?
was (Author: pkolaczk):
I like the idea of using parallelism level as a backpressure mechanism. That
would have a nice positive effect of automatically reducing the amount of
memory used for queuing the requests.
However, I have a few concerns:
# my biggest concern is, that even limiting a single client to one write at a
time (window size = 1), might still be too fast, for some fast clients, if only
row sizes are big enough, particularly when writing big cells of data, where
big = hundreds of kB / single MBs per cell. Cassandra is extremely efficient at
ingesting data into memtables. If it was faster than we're able to flush, then
we still have a problem.
# when not using TokenAware LBP, we can have a problem that in case of
unbalanced write load, the fast (not loaded) nodes would be needlessly
penalized with lower parallelism level, because the window size would be
per-coordinator, not per partition (replica set). Am I right here?
I wonder if using a penalty delay *after* processing the write request, would
not be a better idea: in case of an unbalanced load, when one partition gets
hammered, but most others are ok, it would slow down writes for that one
partition (replica set), but would not affect latency of writes of other
partitions. I'm for applyting the delay after, because then we already know the
replica set and their load, as well as we don't need to keep data queued in
memory, for it has already been written.
The (simplified) process would look as follows:
# the write request gets accepted by the coordinator
# the write gets sent to proper replica nodes
# the replicas that acknowledge the write, also reply with their current load
information
# received load information gets averaged (median)
# when the CL is satisfied, but the load was high enough to be in the
"dangerous zone" the coordinator puts some artificial delay before
acknowledging the write to the client - of course small enough to not exceed
the write timeout. Or if not possible / wise for some other reasons (e.g.
holding memory), just tells the driver that "load was high, you better slow
down" and the driver waits before processing the next request.
WDYT?
> Apply backpressure gently when overloaded with writes
> -----------------------------------------------------
>
> Key: CASSANDRA-7937
> URL: https://issues.apache.org/jira/browse/CASSANDRA-7937
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Environment: Cassandra 2.0
> Reporter: Piotr Kołaczkowski
> Labels: performance
>
> When writing huge amounts of data into C* cluster from analytic tools like
> Hadoop or Apache Spark, we can see that often C* can't keep up with the load.
> This is because analytic tools typically write data "as fast as they can" in
> parallel, from many nodes and they are not artificially rate-limited, so C*
> is the bottleneck here. Also, increasing the number of nodes doesn't really
> help, because in a collocated setup this also increases number of
> Hadoop/Spark nodes (writers) and although possible write performance is
> higher, the problem still remains.
> We observe the following behavior:
> 1. data is ingested at an extreme fast pace into memtables and flush queue
> fills up
> 2. the available memory limit for memtables is reached and writes are no
> longer accepted
> 3. the application gets hit by "write timeout", and retries repeatedly, in
> vain
> 4. after several failed attempts to write, the job gets aborted
> Desired behaviour:
> 1. data is ingested at an extreme fast pace into memtables and flush queue
> fills up
> 2. after exceeding some memtable "fill threshold", C* applies adaptive rate
> limiting to writes - the more the buffers are filled-up, the less writes/s
> are accepted, however writes still occur within the write timeout.
> 3. thanks to slowed down data ingestion, now flush can finish before all the
> memory gets used
> Of course the details how rate limiting could be done are up for a discussion.
> It may be also worth considering putting such logic into the driver, not C*
> core, but then C* needs to expose at least the following information to the
> driver, so we could calculate the desired maximum data rate:
> 1. current amount of memory available for writes before they would completely
> block
> 2. total amount of data queued to be flushed and flush progress (amount of
> data to flush remaining for the memtable currently being flushed)
> 3. average flush write speed
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
