I'd prefer to do MERGE operation last because in H2 it is not standard ANSI SQL MERGE. Or may be not implement it at all, or may be contribute ANSI correct version to H2, then implement it on Ignite. Need to investigate the semantics deeper before making any decisions here.
Lets start with simple scenarios for INSERT and go through all the possible cases and answer the questions: - What will happen on key conflict in TX cache? - What will happen on key conflict in Atomic cache? - What will happen with the previous two if we use DataLoader? - How to make these operations efficient (it will be simple enough to implement them with separate put/putIfAbsent operations but probably we will need some batching like putAllIfAbsent for efficiency)? As for API, we still will need to have a single entry point for all SQL queries/commands to allow any console work with it transparently. It would be great if we will be able to come up with something consistent with this idea on public API. Sergi On Wed, Jul 20, 2016 at 2:23 PM, Dmitriy Setrakyan <dsetrak...@gridgain.com> wrote: > Like the idea of merge and insert. I need more time to think about the API > changes. > > Sergi, what do you think? > > Dmitriy > > > > On Jul 20, 2016, at 12:36 PM, Alexander Paschenko < > alexander.a.pasche...@gmail.com> wrote: > > >> Thus, I suggest that we implement MERGE as a separate operation backed > by putIfAbsent operation, while INSERT will be implemented via put. > > > > Sorry, of course I meant that MERGE has to be put-based, while INSERT > > has to be putIfAbsent-based. > > > > 2016-07-20 12:30 GMT+03:00 Alexander Paschenko > > <alexander.a.pasche...@gmail.com>: > >> Hell Igniters, > >> > >> In this thread I would like to share and discuss some thoughts on DML > >> operations' implementation, so let's start and keep it here. Everyone > >> is of course welcome to share their suggestions. > >> > >> For starters, I was thinking about semantics of INSERT. In traditional > >> RDBMSs, INSERT works only for records whose primary keys don't > >> conflict with those of records that are already persistent - you can't > >> try to insert the same key more than once because you'll get an error. > >> However, semantics of cache put is obviously different - it does not > >> have anything about duplicate keys, it just quietly updates values in > >> case of keys' duplication. Still, cache has putIfAbsent operation that > >> is closer to traditional notion of INSERT, and H2's SQL dialect has > >> MERGE operation which corresponds to semantics of cache put. Thus, I > >> suggest that we implement MERGE as a separate operation backed by > >> putIfAbsent operation, while INSERT will be implemented via put. > >> > >> And one more, probably more important thing: I suggest that we create > >> separate class Update and corresponding operation update() in > >> IgniteCache. The reasons are as follows: > >> > >> - Query bears some flags that are clearly redundant for Update (page > >> size, locality) > >> - query() method in IgniteCache (one that accepts Query) and query() > >> methods in GridQueryIndexing return iterators. So, if we strive to > >> leave interfaces unchanged, we still will introduce some design > >> ugliness like query methods returning empty iterators for certain > >> queries, and/or query flags that indicate whether it's an update query > >> or not, etc. > >> - If some Queries are update queries, then continuous queries can't be > >> based on them - more design-wise ugly checks and stuff like that. > >> - I'm pretty sure there's more I don't know about. > >> > >> Comments and suggestions are welcome. Sergi Vladykin, Dmitry > >> Setrakyan, your opinions are of particular interest, please advise. > >> > >> Regards, > >> Alex >
