Each "per-segment" M task is executed and is offered the way to push
intermediary results in a temp cache.

Just to be clear, the user-provided mapper and combiner don't know anything about the intermediary cache (which doesn't have to be temporary, if it's shared by all M/R tasks). They only interact with the Collector interface.
The map/combine task on the other hand is our code, and it deals with the intermediary cache directly.

Interesting, Evangelos, do you actually use the collector interface or actual explicit intermediary caches in your approach.

If that’s the collector interface, I guess that’s easier to hide that sharding business.

The intermediary results are stored with a composite key [imtermKey-i, seg-j].
The M/R coordinator waits for all M tasks to return. If one does not
(timeout, rehash), the following happens:

We can't allow time out map tasks, or they will keep writing to the intermediate cache in parallel with the retried tasks. So the originator has to wait for a response from each node to which it sent a map task.

OK. I guess the originator can see that a node is out of the cluster though and act accordingly.

- delete [intermKey-i, seg-i] (that operation could be handled by the
new per-segment M before the map task is effectively started)
- ship the M task for that segment-i to the new primary owner of
segment-i

When all M tasks are received the Reduce phase will read all [intermKey-i, *]
keys and reduce them.
Note that if the reduction phase is itself distributed, we could apply
the same key per segment and shipping split for these.

Sure, we have to retry reduce tasks when the primary owner changes, and it makes sense to retry as little as possible.

Again the tricky part is to expose the ability to write to intermediary
caches per segment without exposing segments per se as well as let
someone see a concatenated view if intermKey-i from all segments subkeys
during reduction.

Writing to and reading from the intermediate cache is already abstracted from user code (in the Mapper and Reducer interfaces). So we don't need to worry about exposing extra details to the user.

Thoughts?

Dan, I did not quite get what alternative approach you wanted to
propose. Care to respin it for a slow brain? :)

I think where we differ is that I don't think user code needs to know about how we store the intermediate values and what we retry, as long as their mappers/combiners/reducers don't have side effects.

Right but my understanding from the LEADS guys was that they had side effects on their M/Rs. Waiting for Evangelos to speak up.

Otherwise I was thinking on the same lines: send 1 map/combine task for each segment (maybe with a cap on the number of segments being processed at the same time on each node), split the intermediate values per input segment, cancel+retry each map task if the topology changes and the executing node is no longer an owner. If the reduce phase is distributed, run 1 reduce task per segment as well, and cancel+retry the reduce task if the executing node is no longer an owner.

I had some ideas about assigning each map/combine phase a UUID and making the intermediate keys [intermKey, seg, mctask] to allow the originator to retry a map/combine task without waiting for the previous one to finish, but I don't think I mentioned that before :)

Nice touch, that fixes the rogue node / timeout problem.

There are also some details that I'm worried about:

1) If the reduce phase is distributed, and the intermediate cache is non-transactional, any topology change in the intermediate cache will require us to retry all the map/combine tasks that were running at the time on any node (even if some nodes did not detect the topology change yet). So it would make sense to limit the number of map/combine tasks that are processed at one time, in order to limit the amount of tasks we retry (OR require the intermediate cache to be transactional).

I am not fully following that. What matters in the end it seems is for the originator to detect a topology change and discard things accordingly, no? If the other nodes are slaves of that originator for the purpose of that M/R, we are good.

2) Running a separate map/combine task for each segment is not really an option until we implement the the segment-aware data container and cache stores. Without that change, it will make everything much slower, because of all the extra iterations for each segment.

See my other email about physically merging down the per segment work into a per node work when you ship that work.

3) And finally, all this will be overkill when the input cache is small, and the time needed to process the data is comparable to the time needed to send all those extra RPCs.

So I'm thinking it might be better to adopt Vladimir's suggestion to retry everything if we detect a topology change in the input and/or intermediate cache at the end of the M/R task, at least in the first phase.

You half lost but I think that with my proposal to physically merge the RPC calls per node instead of per segment, that problem would be alleviated.

Emmanuel