Proposal for hash based aggregation in map (in-map combiner)
Pig does (sort based) partial aggregation in map side, through the use of combiner. MR serializes the output of map to a buffer, sorts it on the keys, deserializes and passes the values grouped on the keys to combiner phase. The same work of combiner can be done in the map phase itself by using a hash-map on the keys. This hash based (partial) aggregation can be done with or without a combiner phase.
- It will send fewer records to combiner and thereby -
- Save on cost of serializing and de-serializing
- Save on cost of lock calls on the combiner input buffer. (I have found this to be a significant cost for a query that was doing multiple group-by's in a single MR job. -Thejas)
- The problem of running out of memory in reduce side, for queries like COUNT(distinct col) can be avoided. The OOM issue happens because very large records get created after the combiner run on merged reduce input. In case of combiner, you have no way of telling MR not to combine records in reduce side. The workaround is to disable combiner completely, and the opportunity to reduce map output size is lost.
- When the foreach after group-by has both algebraic and non-algebraic functions, or if a bag is being projected, the combiner is not used. This is because the data size reduction in typical cases are not significant enough to justify the additional (de)serialization costs. But hash based aggregation can be used in such cases as well.
- It is possible to turn off the in-map combine automatically if there is not enough 'combination' that is taking place to justify the overhead of the in-map combiner. (Idea borrowed from Hive.)
- It is not clear, if both MR combiner and in-map combiner should be enabled in a pig MR job. If in-map combiner is used, the data reduction that would happen because of MR combiner might not be sufficient to justify the costs. But MR combiner might help in further reducing the data that gets written to disk, if multiple spill files get merged or if multiple waves of sort-merge happen on reduce side.
- Memory used by the in-map combiner will have to be managed carefully, to avoid any out of memory errors.
Design Option 1
In this option, the work done by MR to group values by key before invoking the combiner plan is simulated within the map using a hash-map. When the hash-map is big enough, run the accumulated groups through the combiner plan. The results of the in-map combine plan execution get written as the map output. This option is easier to implement. But it is not going to be efficient in its memory footprint as input tuples will be kept around until the configured memory limit forces it to combine them. This will result in a smaller number of key-values that will be in memory at a time, and result in fewer values being aggregated, and a larger map output size.
MapReduceOper class that represents MR job in pig MR plan will now have a new member inMapCombinePlan, which is a PhysicalPlan. In the initial implementation, combiner physical plan (the member called combinePlan) can be cloned here.
But for supporting in-map combine for cases where combiner does not get used (eg. when there is a bag/non-algebraic udf), the MR plan optimizer rules need to change. In such cases, the output type of map and combine plan would be different, that could be a problem.
A new class that extends PigMapBase will have a collect call that collects the key-values into a hash-map. The hash-map will spill into the combine plan, when its estimated size exceeds a configurable threshold . This would be similar to the InternalCacheBag implementation.
Design Option 2
In this option, there will be a new physical operator, POGroupHash, that will do the hash based aggregation. This will be the last node in the map plan of MR job corresponding to the group operation. When there are two, or a small number of values for a group key (set), it will compute the new partial aggregate value and store it in the hash-map. The memory management of the hash-map will be similar to that of InternalCacheBag, it will estimate its memory footprint. It will flush some % (5%?) of entries when it exceeds configurable memory limit. The least recently used keys can be chosen to be flushed. THese flushed entries will be written as map output.
It might be useful to have a new udf interface that accepts a tuple at a time to compute a partial aggregate, so that new bags don't have to be created for each new tuple that needs to be aggregated. But the bag creation overhead and overhead of calling the udf multiple times could be reduced by calling the udf only after few values have been accumulated in the hash-map.
This design option will have a smaller memory footprint because input tuples can be aggregated sooner. This will result also in smaller output size because more records can be held in the hash-map. But the work involved with this option is more because the MR plan generation will need to change to use this new relational operator when the query is a 'combinable' query. It can have impact on other visitors and optimizer rules as well.