FWIW, I think the proposed approach to configuration is fine. I think selecting a choice for the user should be done simply and deterministically. We should probably default to Trie based memtables for users with a fresh config file, and we can consider changing the default in a later release for those with an old config file that does not specify an implementation.
From: Dinesh Joshi <djo...@apache.org> Date: Tuesday, 8 February 2022 at 20:21 To: dev@cassandra.apache.org <dev@cassandra.apache.org> Subject: Re: [DISCUSS] CEP-19: Trie memtable implementation Thank you for sharing the perf test results. Going back to the schema vs yaml configuration. I am concerned users may pick the wrong implementation for their use-case. Is there any chance for us to automatically pick a MemTable implementation based on heuristics? Do we foresee users ever picking the existing SkipList implementation over the Trie Given the performance tests, it seems the Trie implementation is the clear winner. To be clear, I am not suggesting we remove the existing implementation. I am for maintaining a pluggable API for various components. Dinesh On Feb 7, 2022, at 8:39 AM, Branimir Lambov <blam...@apache.org<mailto:blam...@apache.org>> wrote: Added some performance results to the ticket: https://issues.apache.org/jira/browse/CASSANDRA-17240 Regards, Branimir On Sat, Feb 5, 2022 at 10:59 PM Dinesh Joshi <djo...@apache.org<mailto:djo...@apache.org>> wrote: This is excellent. Thanks for opening up this CEP. It would be great to get some stats around GC allocation rate / memory pressure, read & write latencies, etc. compared to existing implementation. Dinesh On Jan 18, 2022, at 2:13 AM, Branimir Lambov <blam...@apache.org<mailto:blam...@apache.org>> wrote: The memtable pluggability API (CEP-11) is per-table to enable memtable selection that suits specific workflows. It also makes full sense to permit per-node configuration, both to be able to modify the configuration to suit heterogeneous deployments better, as well as to test changes for improvements such as this one. Recognizing this, the patch comes with a modification to the API<https://github.com/blambov/cassandra/commit/24b558ba2f71a2f040804e28993cc914b31298f5> that defines memtable templates in cassandra.yaml (i.e. per node) and allows the schema to select a template (in addition to being able to specify the full memtable configuration). One could use this e.g. by adding: memtable_templates: trie: class: TrieMemtable shards: 16 skiplist: class: SkipListMemtable memtable: template: skiplist (which defines two templates and specifies the default memtable implementation to use) to cassandra.yaml and specifying WITH memtable = {'template' : 'trie'} in the table schema. I intend to commit this modification with the memtable API (CASSANDRA-17034/CEP-11). Performance comparisons will be published soon. Regards, Branimir On Fri, Jan 14, 2022 at 4:15 PM Jeff Jirsa <jji...@gmail.com<mailto:jji...@gmail.com>> wrote: Sounds like a great addition Can you share some of the details around gc and latency improvements you’ve observed with the list? Any specific reason the confirmation is through schema vs yaml? Presumably it’s so a user can test per table, but this changes every host in a cluster, so the impact of a bug/regression is much higher. On Jan 10, 2022, at 1:30 AM, Branimir Lambov <blam...@apache.org<mailto:blam...@apache.org>> wrote: We would like to contribute our TrieMemtable to Cassandra. https://cwiki.apache.org/confluence/display/CASSANDRA/CEP-19%3A+Trie+memtable+implementation This is a new memtable solution aimed to replace the legacy implementation, developed with the following objectives: - lowering the on-heap complexity and the ability to store memtable indexing structures off-heap, - leveraging byte order and a trie structure to lower the memory footprint and improve mutation and lookup performance. The new memtable relies on CASSANDRA-6936 to translate to and from byte-ordered representations of types, and CASSANDRA-17034 / CEP-11 to plug into Cassandra. The memtable is built on multiple shards of custom in-memory single-writer multiple-reader tries, whose implementation uses a combination of state-of-the-art and novel features for greater efficiency. The CEP's JIRA ticket (https://issues.apache.org/jira/browse/CASSANDRA-17240) contains the initial version of the implementation. In its current form it achieves much better garbage collection latency, significantly bigger data sizes between flushes for the same memory allocation, as well as drastically increased write throughput, and we expect the memory and garbage collection improvements to go much further with upcoming improvements to the solution. I am interested in hearing your thoughts on the proposal. Regards, Branimir
