Re: [infinispan-dev] Major version cleaning
On 02/21/2017 05:16 PM, Tristan Tarrant wrote:
On 21/02/17 16:29, Sanne Grinovero wrote:
>> You haven't explained what "flush" means. Since you separate that
from
>> atomicity/consistency, I assume that batches on non-tx cache are just
>> ordered putOrRemoveAll operations, immediately visible on flush without
>> any atomicity.
I assume that in Sanne's idea, ordering in a batch doesn't matter, aside
from operations on the same key. Having ordering in there would for
example not allow us to parallelize by segment.
> So I want to write a first chunk, in our code that looks like:
>
> startBatch
> put(chunk1/A, [some large value])
> put(chunk1/B, [some small metadata])
> put(chunk1/C, [some small metadata])
> endBatch
> There is no reason to use a transaction, in fact we had to disable
> transactions as some of these entries could be large.
> There also is no reason for the batch, other than optimising the latency.
Let me summarize to see if we have the requirements for a useful
batching system (which is sort of patterned on the JDBC statement batching):
- a batch is not an atomic operation, i.e. it is not backed by a transaction
- it can be wrapped in a transaction if needed
- batches cannot be nested
- batches only involve unconditional write operations (put, putAll, remove)
- ordering of operations within a batch is unimportant aside from
modifications to the same key where we apply "last one wins"
- when a batch is "flushed" (i.e. endBatch is invoked) the ops are
grouped by segment and sent to the appropriate owner for processing,
potentially in parallel
As Radim has called it, this is essentially a putOrRemoveAll op (with an
async counterpart).
It is putOrRemoveAll when applied on a non-tx cache, and actually
implementing that shouldn't be complex. However, when transactions come
into play, it is different, because Sanne wants us to remove the
modifications in completed batch from the local transactional invocation
context and 'cache' them on the owners. Since reads have to be
transactionally consistent, we need to inspect the transaction on the
remote nodes (remote repeatable read).
Sanne's request makes sense to me. However as the current implementation
is providing a false assumption that it could work as JDBC batches while
it's nothing but crippled JTA, and as I don't see anyone shouting "I'll
implement that, next week it's done!", I second deprecating/removing the
API for the time being.
I don't find the current API ideal either, as it depends on thread
locals (JTA does as well, but...) while it does not seem useful enough
to me. I would prefer
interface BatchingCache {
Batch start();
}
@NotThreadSafe
interface Batch {
void put(K k, V value);
...
void execute();
void drop(); // maybe not needed
}
Radim
Is that summary correct ?
Tristan
--
Radim Vansa <rvansa(a)redhat.com>
JBoss Performance Team