The Container is the home for all global processing of user actions (represented as queries) and their results. It provides a development and hosting environment for processing components, and a model for composing such components developed by multiple development teams into a functional whole.
This document describes how to develop and deploy Searcher components. To get started with development, see the Developer Guide. For reference, see the Container javadoc, and the services.xml reference.
Best practise for queries is submitting the user-generated query as-is to Vespa, then use Searcher components to implement additional logic. Refer to the Query HTTP API.
The components of the search container are called Searchers.
A searcher is a component - usually deployed as part of an OSGi bundle -
which extends the class
All Searchers must implement a single method:
When the container receives a request, it will create a Query representing it
and execute a configured list of such Searcher components.
This is done by calling the
search() method on the first searcher in the list.
That searcher is responsible for passing the call to the next searcher in the list (or not, as it sees fit).
This is done by calling
search() on the Executor given,
which keeps track of where we are in the list of Searchers.
Hence, this is a noop searcher implementation:
Eventually the search call will reach the end of the list of searchers. The last searcher in the list may create a Result (somehow), which is now passed back up the call chain until it reaches the top. The container will then translate that Result back to a response to the incoming request.
As is evident from this description, this is a synchronous model, where each request is processed in a dedicated worker thread until the result is returned. This synchronous model is implemented with multi-threading of individual searchers.
The single searcher method is sufficient to express all kinds of functionality, e.g.:
The Query in the search container is the container of all the information needed to create a result to the request, including:
The Result encapsulates all the data generated from a Query. The Result contains a composite tree of Hit objects organized in lists called HitGroups (the Result points to the topmost group). Each Hit contains some particular data item which is deemed relevant to the Query. The Hit objects has a general key-value storage, but are also polymorphic to support representing more structured information. See the inspecting structured data documentation for details about handling structured information in a Searcher.
As Hits may be hierarchically organized into hit lists, the Result object is capable of representing any organization of the results. For example, in a federated system the hits are initially organized in one hit group per source. Upstream searchers may reorganize this into something that fits the user's need better, e.g a single blended group, or one group per likely interpretation of the query etc.
The lists of searchers mentioned above are called search chains. Search chains are a special case of the general component chains. A search chain is nothing more than a list of searcher instances having an id. The search chains are typically not created for every query but are part of the configuration. Multiple ones may exist at the same time, the chain to execute may be specified in the request. If nothing is specified, a default one is used. The same Searcher instance may exist in multiple search chains, which is why the Execution object is responsible for knowing the next Searcher to invoke in a particular request.
Search chains may also be executed programmatically (typically from a Searcher), synchronously or asynchronously:
Asynchronous execution is useful in cases like federation, where a searcher forks a Query to multiple search chains in parallel, each getting results from a particular source. Also, as in the example, it is allowed to use the same Execution instance to construct multiple AsyncExecution instances, as the state is only copied from the constructor argument.
The execution order of the searchers in a chain are not ordered explicitly, but by ordering constraints declared in the searchers or their configuration. Also read the search reference.
Example of a complete searcher:
The container will create one or more instances of this class and place it in the desired search chain(s) to serve queries, as specified in the configuration. The first line in this searcher forwards the query to whatever is the next searcher in the chain this is a part of. This will eventually produce a Result, which is modified and then passed back to the previous searcher in this chain. The container will create a new instance of this searcher only when it is reconfigured, so any data needed by the searcher can be read and prepared from a constructor in the searcher. Constructors may also accept configuration, as any other pluggable component.
Find the full API available to searchers in the Search Container Javadoc.
Before there is any point in testing a searcher in a real system, it should pass a set of unit tests which test it in isolation or together with the few searchers it interacts with. To do this, we can write unit tests which programmatically sets up a search chain containing the searcher to be tested, the searchers it interacts with (if any) and a searcher which produces mock results appropriate for the tests. Here is a simple example testing the Searcher above:
In this case, no searcher producing mock results is needed because the searcher we are testing does not care what the Result contains. If the search chain ends with a searcher which produces no result, the framework will simply return an empty result, which is what happens in this case. A test adding a mock searcher producing results are shown in federation.
To write unit tests of the whole application package, see the Developer Guide.
Once the searcher passes unit tests, it can be deployed to the Vespa system hosting it. The procedure is the same as described in deploying a component. First build the component jar. To include the searcher in services.xml, define a search chain and add the searcher to it - example:
This defines the search chain
which will be used in queries when no other chain is explicitly specified.
The searcher id above is resolved to the component bundle jar we added by the symbolic name in the manifest,
and to the right class within the bundle by the class name.
By keeping all these three the same, we keep things simple,
but more advanced use where this is possible is also supported, see later sections.
See the search chains reference.
By creating a directory containing this services.xml, hosts.xml and components/Simplesearcher.jar, that directory becomes a complete application package containing a bundle, which can now be deployed to a Vespa instance.
After deployment, query the application: http://localhost:8080/search?query=best.
A searcher can also be tested running inside a container. Create an instance from the container part of the services.xml file above:
Examine which searchers are in a chain and their ordering:
The query object is used to pass information between searchers. A part of the query is a general property store which may hold any object. Any values set in the request or in the query profile is available through these properties, but in addition searchers may add any objects they create. This is useful when some searcher component is producing information later consumed by some other. Example:
This code illustrates two idioms such searchers should follow when exchanging data:
When it does not cause unwanted dependencies,
it is recommended to wrap the property get and put in a (static)
setTo method in the stored object,
to allow storage and lookup without having to mention the key unnecessarily outside the object.
Note that the objects are passed as regular in-memory references, so there is no noticeable overhead in this. However, in some situations (like when federating to multiple sources) the query will need to be cloned. The query will then attempt to clone the added properties. Those that implement Cloneable will have clone called, the rest will be copied by reference.
On the other hand, cloning objects which should not change is wasteful, they should be copied by reference. Hence, the guidelines are:
In some cases there is a need for passing information between searchers beyond those who see the same Query object. For this purpose, the Query provides a QueryContext object which provides a shared data view to all Searchers working on the same request. The context provides (among other things) a facility for setting properties (named objects). The context can be accessed safely from all the threads working on a request without incurring synchronization overhead (with some caveats), but provides linear, not constant lookup time. To set and retrieve such properties, use:
It is easy to pass arguments to searchers - any key-value looked up in the query properties in the searcher can be passed as is in the request, or in a query profile. Example:
This value can be set by adding
&my.parameter=myValue to the request. Guidelines:
myfeature.betc. In addition to helping keep the search API clean, this allows various query profiles containing settings for all values in
myfeatureto be defined and selected at run time, which is often useful.
Parameters should be used for all query state which it is reasonable and just as cheap to assume may change with every query. Good candidates are e.g. numerical values to algorithms and switches to business logic.
In broad strokes, the Container works like this:
There is a single instance of each search chain. In every chain, there is a single instance of each searcher. (Unless a chain is configured with multiple, identical searchers - this is a rare case.)
When simultaneous requests arrive for the same search chain,
multiple worker threads execute the searchers in that chain.
A searcher can therefore be executed concurrently by multiple threads;
many threads of execution can be going through the
search() method, concurrently.
This model places an important constraint on searcher classes: instance variables are not safe. They must be eliminated, or made thread-safe somehow.
As the passing of queries and results happen on the call stack, the container will allocate many worker threads to execute queries, using one thread per query until the result is returned.
This means that any state we wish to keep along in the searcher for this particular query until the result is returned should be kept as local variables in the search method, while state which should be shared by all queries should be kept as member variables. As the latter kind will be accessed by multiple threads at any one time, the state of such member variables must be multithread safe.
This critical restriction is similar to those of e.g. the Servlet API. A quick example should drive the point home:
The second example uses an instance variable, which will be accessed concurrently by multiple threads. Without proper concurrency controls (such as synchronization), such access is inherently unsafe and may yield inconsistent results, and/or data corruption.
Options for implementing a multithread-safe searcher with instance variables:
The model of a single pass fetching results from a Query described in this document is sometimes too simplistic to produce good performance. The search container supports multiphase searching to address such cases. With multiphase searching, the hits of the result is first filled with some minimal information. This minimally filled result is sent up the search chain where some of the hits are hopefully removed. When more information is needed, a second fill request is sent down the search chain to fetch more data for just those hits remaining in the result. This can happen in repeated stages, working on progressively smaller sets of hits containing progressively more expensive information.
The container supports this by offering
fill methods on execution,
which may be called to request more information added to the hits of the result from a searcher.
In addition, the backends and backend providers must support multiphase searching
(this is currently only the case for internal Vespa clusters).
Any searchers should assume they are operating in a multi-phase setup, meaning:
fill()to get the default set of fields for each hit type or
fill(summaryClassName)to get a particular collection of fields known to exist in the backend(s) in question. Calling fill on a result which contains already-filled hits is cheap.
searchmethod and the
fillmethod. The fill method must request filling down the source branches which has remaining hits in the result.
fillmethod which fills the hits in the given result belonging to that backend with information from the backend.
If your searcher encounters a problem and wants to signal an error,
set an error hit in the result object by calling
See the FAQ for timeouts.
In a Searcher, one often will use
to modify the current query, or create a new query based on input query terms, or results from the current query.
To keep linguistic settings (e.g. stemming) from the parent query, set
isFromQuery to true -