Thursday, February 3, 2011

Mondrian SPI SegmentCache

Fellow Mondrian developers and users,

One month has already passed since the new year festivities, and while most of you have been trying to renew your gym membership or hold on to your new year resolutions the best you could, so did the Mondrian team. Our resolutions, although not requiring personal sacrifices, are none the less starting to bear fruit.

For you see, our resolution for the year was to provide Mondrian developers and integrators means to achieve better understanding, scalability and control. We have many ideas on how to reach those goals. Some of them are still in their infancy, yet some of them have already been committed to the source. Last month, we worked on the first phase. We added means for system architects to externalize and share a  pluggable segment cache. What does this mean exactly? Let's take a step back in order to better understand.

Internally, Mondrian splits the tuples in segments. A typical segment could be described as a measure crossjoined by a series of predicates. As an example, a textual representation of a segment contents could be:
Measure = [ Sales ]
Predicates = {
          [ Products = * ],
    [ State = California ],
    [ Gender = Male ] }
Data = [ 1346.34, 234.00, ... ]
In the case above, the segment would represent the Sales data of all males in California, for all products. It is a lot more effective to deal with those data structures. If Mondrian was to internally represent each data cell individually, the unique identifier of that cell would be of a greater size than the data itself, thus creating a whole lot more problems in terms of data efficiency. This is therefore why Mondrian deals with groups of cells, which it loads in batches, rather than individually. There is a lot of voodoo magic and heuristics in the background trying to figure out how best to group those segments and how to reduce the number of segments to load, ultimately reducing the number of SQL queries to be executed. Mondrian will group all segments with the same predicates but with a different measure into a segment group. Mondrian will also tend to remove as many predicates as it possibly can in order to optimize the data payload. Lets say that a segment covers all products except a single one, Mondrian will still include the product in the segment but filter it out when a specific query requires it.

Once those segments are populated, Mondrian keeps those in a collection of weak references in local memory. All required segment references are pinned down during the resolving of a particular query, but as soon as the query is done executing, the references are returned to their weak state, thus ready to be garbage collected if needed. This simple mechanism allows Mondrian to answer just about any query, as long as the memory allocated is big enough to answer that particular query. This works really well in fact, since in most small deployments, the maximum amount of memory is never reached. And if it ever gets filled, old segments will be evicted to make some room for the new ones.

Now, there are obvious gotchas. First off, what if it takes a long time for a segment to be populated by the RDBMS. This means that if a particular segment ever gets picked up by the garbage collector, the MDX query sent to Mondrian *might* take longer to execute, whether it was in the segment cache or not. This is not acceptable, simply because this makes all performance predictions impossible.

This is where the SegmentCache SPI comes in. It is essentially a pluggable cache for segments. The algorithm behind the segment loader becomes this:
  • Lookup segments in local cache and pin those required.
  • Optimize / group segments
  • Lookup segments from the SPI cache
  • Load the segments found from the SPI  cache
  • Populate the remaining unloaded segments from the RDBMS
  • Put the segments which come from the RDBMS into the SPI cache
  • Pin all loaded segments
  • Resolve the query
  • Unpin all segments in the local cache
But wait! There is more! The SegmentCache SPI is trivial to implement.
Future<Boolean> contains(SegmentHeader header);
Future<SegmentBody> get(SegmentHeader header);
Future<List<SegmentHeader>> getSegmentHeaders();

Future<Boolean> put(
  SegmentHeader header,
  SegmentBody body);
void tearDown();
Figure 1. Mondrian Segment Loader Architecture


There are two assumptions that are made towards the implementation. The first obvious one is that the cache must assume that many Mondrian instances might access the cache concurrently, form different threads. We therefore recommend using the Actor Pattern or anything similar in order to enforce thread safety. The second is that SegmentCache implementations will be instantiated very often. We therefore recommend using a facade object which relays calls to the actual segment cache code. Update: This was redesigned so that a singleton is created and used throughout Mondrian's internals.

As for the storage of the SegmentHeader and SegmentBody objects, we tried to make it as simple and flexible as possible. Both objects are fully serializable and are immutable. They are also specially crafted to use dense arrays of primitive data types. We also tried to make extensive use of Java native functions when copying the data to / from the cache within Mondrian internals.

The bottom line is that from now on the Mondrian community will be free to implement segment caches to fit their needs. We will be rolling out a few default implementations and examples, obviously. One neat implementation could be one which pages the segments to a super fast array of SSD drives. Another one could be to store the segments in Terracota or ehCache or Infinispan, or just about any scalable caching system there is out there. So if any of you out there are interested in implementing this SPI for your business and would like to either share your experiences or contribute those implementations, don't hesitate to contact us. Or me directly.

There is more goodness to come, but that's it for now. Stay tuned!

3 comments:

  1. Luc,

    this is cool stuff!

    I really enjoy reading these technical blogs from you, keep it up :)

    ReplyDelete
  2. Hi Luc,

    Now that mondrian 3.3 is out, we were trying to use the SPI and implement external cache.

    If I have understood your post correctly, this new interface is something like a fallback and runs parallel to the default implementation that uses JVM memory.

    However we could not understand the algorithm behind the segment loader. It would be great if you can give some more insights or examples on the same.

    Regards
    Altaf

    ReplyDelete
    Replies
    1. We have substantially refactored the way Mondrian internally handles the segments. As soon as 3.4.0 is out of the door, we will be explaining the new approach in great details. Stay tuned!

      Delete