Custom SOLR Search Components - 2 Dev Tricks

I've been building some custom search components for SOLR lately, so wanted to share a couple of things I learned in the process. Most likely this is old hat to people who have been doing this for a while, but thought I'd share, just in case it benefits someone...

Passing State

In a previous post, I described a custom SOLR search handler returns layered search results for a given query term (and optional filters). As I went further, though, I realized that I needed to return information relating to facets and category clusters as well. Of course, I could have added this stuff into the handler itself, but splitting the logic across a chain of search components seemed to be more preferable, readability and reusability wise, so I went that route.

So the first step was to refactor my custom SearchHandler into a SearchComponent. Not much to do there, except to subclass SearchComponent instead of RequestHandlerBase and move the handleRequestBody(SolrQueryRequest,SolrQueryResponse) to a process(ResponseBuilder) method. The request and response objects are accessible from the ResponseBuilder as properties, ie, ResponseBuilder.req and ResponseBuilder.rsp. I then declared this component and an enclosing handler in solrconfig.xml, something like this:

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26

<!-- this used to be my search handler --> <searchComponent name="component1" class="org.apache.solr.handler.component.ext.MyComponent1"> <str name="prop1">value1</str> <str name="prop2">value2</str> </searchComponent> <searchComponent name="component2" class="org.apache.solr.handler.component.ext.MyComponent2"> <lst name="facets"> <str name="prop1">1</str> <str name="prop2">2</str> </lst> </searchComponent> <requestHandler name="/mysearch2" class="org.apache.solr.handler.component.SearchHandler"> <lst name="defaults"> <str name="echoParams">explicit</str> <str name="fl">*,score,id</str> <str name="wt">xml</str> </lst> <arr name="components"> <str>component1</str> <str>component2</str> <!-- ... more components as needed ... --> </arr> </requestHandler>

I've also added a second component to the chain above (just so I don't have to show this snippet again later), hope its not too confusing. Obviously there can be multiple components before and after my search handler turned search component, but for the purposes of this discussion, I'll keep things simple and just concentrate on this one other component and pretend that it has multiple unique (and pertinent) requirements.

Now, assume that the second component needed data that was already available, or can be easily generated by component1. Its actually true in my case, since I needed a BitSet of document ids in the search results in my second component, which I could easily get by collecting them while looping through the SolrDocumentList of results in my first component. So it seemed kind of wasteful to compute this again. So I updated this snippet of code in component1's process() method (what used to be my handleRequestBody() method):

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

public void process(ResponseBuilder rb) throws IOException { ... // build and write response ... OpenBitSet bits = new OpenBitSet(searcher.maxDoc()); List<SolrDocument> slice = new ArrayList<SolrDocument>(); for (Iterator<SolrDocument> it = results.iterator(); it.hasNext(); ) { SolrDocument sdoc = it.next(); ... bits.set(Long.valueOf((Integer) sdoc.get("id"))); if (numFound >= start && numFound < start + rows) { slice.add(sdoc); } numFound++; } ... rsp.add("response", results); rsp.add("_bits", bits); }

In my next component (component2), I simply grab the OpenBitSet data structure by name from the NamedList, use them to generate the result for this component, stick the result back into the response, and discard the temporary data. The last is so that the data does not appear on the response XML (for both aesthetic and performance reasons).

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

public void process(ResponseBuilder rb) throws IOException { Map<String,Object> cres = new HashMap<String,Object>(); NamedList nl = rb.rsp.getValues(); OpenBitSet bits = (OpenBitSet) nl.get("_bits"); if (bits == null) { logger.warn("Component 1 must write _bits into response"); rb.rsp.add(COMPONENT_NAME, cres); return; } // do something with bits and generate component response doSomething(bits, cres); // stick the result into the response and delete temp data rb.rsp.add("component2_result", cres); rb.rsp.getValues().remove("_bits"); }

Before I did this, I investigated if I could subclass the XmlResponseWriter to ignore NamedLists with "hidden" names (ie names prefixed with underscore), but the XmlResponseWriter calls XMLWriter which does the actual XML generation, and XMLWriter is final (at least in SOLR 1.4.1). Good thing too, forced me to look for and find a simpler solution :-).

So there you have it - a simple way to pass data between components in a SOLR Search RequestHandler. Note that it does mean that component2 is always dependent on component1 (or some other component that produces the same data) upstream to it, so these components are no longer truly reusable pieces of code. But this can be useful if you really need it and you document the requirement (or complain about it if not met, as I've done here).

Reacting to a COMMIT

The second thing I needed to do in component2 was to give it some reference data that it would need to compute its results. The reference data is generated from the contents of the index, and the generation is fairly heavyweight, so you don't want to do this on every request.

Now one of the cool things about SOLR is its built-in incremental indexing feature (one of the main reasons we considered using SOLR in the first place), so you can POST data to a running SOLR instance followed by a COMMIT, and voila: your searcher re-opens with the new data.

Of course, this also means that if we want to provide accurate information, the reference data should be regenerated whenever the searcher is reopened. The way I went about doing this is mostly derived from how the SpellCheckerComponent does it, in order to regenerate its dictionaries -- by hooking into the SOLR event framework.

To do this, my component2 implements SolrCoreAware in addition to extending SearchComponent. This requires me to implement the inform(SolrCore) method, which is invoked by SOLR after the init(NamedList) but before prepare(ResponseBuilder) and process(ResponseBuilder). In the inform(SolrCore) method, I register a listener for the firstSearcher and newSearcher events (described in more detail here).

I then build the inner listener class, which implements SolrEventListener, which requires me to provide implementations for newSearcher() and postCommit() methods. Since my listener is a query-side listener, I provide an empty implementation for postCommit(). The newSearcher() method contains the code to generate the reference sets. Here is the relevant snippet of code from the component.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47

public class MyComponent2 extends SearchComponent implements SolrCoreAware { private RefData refdata; // this needs to be regenerated on COMMIT @Override public void init(NamedList args) { ... } @Override public void inform(SolrCore core) { listener = new MyComponent2Listener(); core.registerFirstSearcherListener(listener); core.registerNewSearcherListener(listener); } @Override public void prepare(ResponseBuilder rb) throws IOException { ... } @Override public void process(ResponseBuilder rb) throws IOException { ... // do something with refdata ... } private class MyComponent2Listener implements SolrEventListener { @Override public void init(NamedList args) { /* NOOP */ } @Override public void newSearcher(SolrIndexSearcher newSearcher, SolrIndexSearcher currentSearcher) { RefData copy = new RefData(); copy = generateRefData(newSearcher); refdata.clear(); refdata.addAll(copy); } @Override public void postCommit() { /* NOOP */ } } ... }

Notice that I have registered the listener to listen on both firstSearcher and newSearcher events. This way, it gets called on SOLR startup (reacting to a firstSearcher event), and again each time the searcher is reopened (reacting to a newSearcher event).

One other thing... since the generation of RefData takes some time, its best to have the listener's newSearcher method build a copy and then repopulate the refdata variable from the copy, that way the component continues to use the old data until the new one is available.

And thats pretty much it for today. Till next time.