For analytics use cases, streaming expressions make it possible to compute basic aggregations (count, min, max, sum, and avg) over massive data sets. Moreover, with massive data sets, it is common to use collection aliases over many underlying collections, for instance time-partitioned aliases backed by a set of collections, each covering a specific time range. In some cases, we can end up with many collections (think 50-60) each with 100's of shards. Aliases help insulate client applications from complex collection topologies on the server side.
Let's take a basic facet expression that computes some useful aggregation metrics:
Behind the scenes, the FacetStream sends a JSON facet request to Solr which then expands the alias to a list of collections. For each collection, the top-level distributed query controller gathers a candidate set of replicas to query and then scatters distrib=false queries to each replica in the list. For instance, if we have 60 collections with 200 shards each, then this results in 12,000 shard requests from the query controller node to the other nodes in the cluster. The requests are sent in an async manner (see SearchHandler and HttpShardHandler) In my testing, we’ve seen cases where we hit 18,000 replicas and these queries don’t always come back in a timely manner. Put simply, this also puts a lot of load on the top-level query controller node in terms of open connections and new object creation.
Instead, we can use plist to send the JSON facet query to each collection in the alias in parallel, which reduces the overhead of each top-level distributed query from 12,000 to 200 in my example above. With this approach, you’ll then need to sort the tuples back from each collection and do a rollup, something like:
One thing to point out is that you can’t just avg. the averages back from each collection in the rollup. It needs to be a weighted avg. when rolling up the avg. from each facet expression in the plist. However, we have the count per collection, so this is doable but will require some changes to the rollup expression to support weighted average.
While this plist approach is doable, it’s a pain for users to have to create the rollup / sort over plist expression for collection aliases. After all, aliases are supposed to hide these types of complexities from client applications!
The point of this ticket is to investigate the feasibility of auto-wrapping the facet expression with a rollup / sort / plist when the collection argument is an alias with multiple collections; other stream sources will be considered after facet is proven out.
Lastly, I also considered an alternative approach of doing a parallel relay on the server side. The idea is similar to plist but instead of this being driven on the client side, the FacetModule can create intermediate queries (I called them relay queries in my impl.) that help distribute the load. In my example above, there would be 60 such relay queries, each sent to a replica for each collection in the alias, which then sends the distrib=false queries to each replica. The relay query response handler collects the facet responses from each replica before sending back to the top-level query controller for final results.
I have this approach working in the attached patch (relay-approach.patch) but it feels a little invasive to the FacetModule (and the distributed query inner workings in general). To me, the auto- plist approach feels like a better layer to add this functionality vs. deeper down in the facet module code. Moreover, we may be able to leverage the plist solution for other stream sources whereas the relay approach required me to change logic in the FacetModule directly, so is likely not as reusable for other types of queries. It's possible the relay approach could be generalized but I'm not clear how useful that would be beyond streaming expression analytics use cases; feedback on this point welcome of course.
I also think plist should try to be clever and avoid sending the top-level (per collection) request to the same node if it can help it.