[
https://issues.apache.org/jira/browse/CASSANDRA-11407?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Anubhav Kale updated CASSANDRA-11407:
-------------------------------------
Description:
Today's DTCS implementation has been discussed and debated in a few JIRAs
already (the notable one is
https://issues.apache.org/jira/browse/CASSANDRA-9666). One of the main
challenges with the current approach is that it is very difficult to reason
about how the "Target" class makes buckets, thus making it difficult to reason
about the expected file layout on disk.
I am proposing a simplification to current approach that keeps most of the DTCS
properties intact that makes it a great fit for time-series data. The
simplification is as follows.
Given the min and max timestamps across all SS Tables in question, start from
min and make windows based on base and min_threshold. The logic in GetWindow
simply tries to fit maximum sized windows from min to max.
This keeps the DTCS properties intact except that we don't need to wait for
min_threshold windows before making a bigger one. I would argue this simplifies
the algorithm to a great extent, is easy to reason about and the end result
isn't drastically different than the original DTCS in most cases. We give up on
the "alignment" logic that exists in current implementation, but I honestly
don't think it buys us a lot besides complexity.
The implementation can obviously be optimized and cleaned up more if folks
think this is a good idea.
was:
Today's DTCS implementation has been discussed and debated in a few JIRAs
already (the notable one is
https://issues.apache.org/jira/browse/CASSANDRA-9666). One of the main
challenges with the current approach is that it is very difficult to reason
about how the "Target" class makes buckets, thus making it difficult to reason
about the expected file layout on disk.
I am proposing a simplification to current approach that keeps most of the DTCS
properties intact that makes it a great fit for time-series data. The
simplification is as follows.
Given the min and max timestamps across all SS Tables in question, start from
min and make windows based on base and min_threshold. The logic in GetWindow
simply tries to fit maximum sized windows from min to max.
This keeps the DTCS properties intact except that we don't need to wait for
min_threshold windows before making a bigger one. I would argue this simplifies
the algorithm to a great extent, is easy to reason about and the end result
isn't drastically different than the original DTCS in most cases. We give up on
the "alignment" logic in current class, but I honestly don't think it buys us a
lot besides complexity.
The implementation can obviously be optimized and cleaned up more if folks
think this is a good idea.
> Proposal for simplified DTCS
> ----------------------------
>
> Key: CASSANDRA-11407
> URL: https://issues.apache.org/jira/browse/CASSANDRA-11407
> Project: Cassandra
> Issue Type: Improvement
> Components: Compaction
> Reporter: Anubhav Kale
> Attachments: 0001-Simple-DTCS.patch
>
>
> Today's DTCS implementation has been discussed and debated in a few JIRAs
> already (the notable one is
> https://issues.apache.org/jira/browse/CASSANDRA-9666). One of the main
> challenges with the current approach is that it is very difficult to reason
> about how the "Target" class makes buckets, thus making it difficult to
> reason about the expected file layout on disk.
> I am proposing a simplification to current approach that keeps most of the
> DTCS properties intact that makes it a great fit for time-series data. The
> simplification is as follows.
> Given the min and max timestamps across all SS Tables in question, start from
> min and make windows based on base and min_threshold. The logic in GetWindow
> simply tries to fit maximum sized windows from min to max.
> This keeps the DTCS properties intact except that we don't need to wait for
> min_threshold windows before making a bigger one. I would argue this
> simplifies the algorithm to a great extent, is easy to reason about and the
> end result isn't drastically different than the original DTCS in most cases.
> We give up on the "alignment" logic that exists in current implementation,
> but I honestly don't think it buys us a lot besides complexity.
> The implementation can obviously be optimized and cleaned up more if folks
> think this is a good idea.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
