As Jacob mentioned there is not direct relationship between the rocksdb tts (internal to rocksdb) and changelog (done by Samza). The problem may arise if the store is restored from the changelog, since the log will have the expired entries, and they will be entered with the NEW date (and as Yi mentioned, there is no TTL on kafka based changelogs now). But since it is not an error per se, SAMZA-862 <https://issues.apache.org/jira/browse/SAMZA-862> has changed this message to be a warning instead of an error.
On Thu, Jan 28, 2016 at 11:51 AM, Yi Pan <nickpa...@gmail.com> wrote: > Hi, David, > > The "compaction" referred to together w/ TTL is referring to RocksDb's > compaction, not the Kafka-based changelog topic. Currently, TTL is not > applied to Kafka-based changelog topic. SAMZA-677 is opened for this. > > -Yi > > On Thu, Jan 28, 2016 at 11:36 AM, David Garcia > <dgar...@homeaway.com.invalid > > wrote: > > > Ok, that makes sense. I had assumed that the changelog was supported > > because the docs mention that TTL is enforced upon ³compaction² (I had > > assumed compaction of the DB changelog). Which topic does the TTL policy > > listen for the compaction of (since compaction policies of topics can > > differ)? > > > > -David > > > > On 1/27/16, 8:46 PM, "Jacob Maes" <jacob.m...@gmail.com> wrote: > > > > >Here's my understanding. The others can correct me if I'm mistaken. > > > > > >Samza provides the changelog functionality by intercepting RocksDB "put" > > >and "delete" operations. However, TTL is managed by RocksDB internally > and > > >there aren't any hooks exposed in the RocksDB JNI. So there are 2 > problems > > >that arise with TTL and change logging: > > >1. Samza doesn't know when an entry expires, so it can't delete the > > >expired > > >entry from the changelog. > > >2. The changelog currently has no concept of entry age/timestamp, so > when > > >the changelog is restored, it's unknown whether some subset (or all) of > > >the > > >entries should be immediately expired. > > > > > >These issues aren't insurmountable, but they weren't pursued for the > > >initial implementation. Perhaps because there was a shortage of use > cases > > >that needed both TTL and changelogging, but I'm not sure. > > > > > >-Jake > > > > > >On Wed, Jan 27, 2016 at 6:19 PM, David Garcia > > ><dgar...@homeaway.com.invalid> > > >wrote: > > > > > >> So, I saw this very scary message: > > >> > > >> > > >> ERROR - e.kv.RocksDbKeyValueStore$ - sessionJoinStore is a TTL based > > >> store, changelog is not supported for TTL based stores, use at your > own > > >> discretion > > >> > > >> > > >> > > >> > > >> A few of questions: > > >> > > >> 1.) Does this mean that this store is NOT backed by the changelog? > > >> > > >> 2.) Provided that the store IS backed by a change log, do the TTL > > >> expirations commit removals from the changelog (I.e. > Nulls)...presumably > > >> upon compaction > > >> > > >> 3.) Can I please get a bit more detail on how TTL affects a changelog > > >> store? > > >> > > >> > > >> -David > > >> > > > > >
