This is the pure Java implementation of the TMAP Live Access Server. There are three major components to this implementation of LAS, the Product Server, a collection of backend services and the user interface server.

The Product Server

The LAS Product Server (LPS) is implemented as a Struts-based Web application server (though care has been taken to keep the logic that deals with LAS configuration files and the request and response messages separate from the framework). The LPS typically does work on behalf of an LAS User Interface (LUI) server, a servlet-based Web application server responsible for presenting HTML pages to a browser from which the user interacts with an LAS site. When the user has completed the selections necessary to create an LAS Product, the LUI redirects the browser to an LPS URL. The query string of this URL contains one essential parameter (a text parameter named "xml" which is a string representation of the UI Request XML [some outdated documentation about these messages is available here: http://www.ferret.noaa.gov/Ferret/LAS/Documentation/manual/reqapi.html). The user interface model we use is for the LPS URL to be opened in a new "output" window of the user's browser. This window accumulates LPS URLs in its history and interacts with the LPS to give the user feedback from long-running product requests, gives the user the ability to cancel long-running product requests and has the potential to make subsequent requests directly to the LPS. The remainder of this document will deal exclusively to what happens after the request has been received by the LPS.

The Life-cycle of a Request

The life-cycle of a LAS UI Request from its arrival at the LPS is as follows. Using the UI Request and the LAS configuration, the LPS (a Struts Action class called ProductServerAction) creates a ProductRequest object. Using the newly constructed ProductRequest object along with the LAS configuration (which has been processed when the servlet initialized to merge the hierarchy of <property> definitions), the LPS specific configuration (information about the cache and the known backend services), the incoming HTTP Request object, the Structs action mapping (this is unnecessary and will be eliminated to separate the "Runner" thread from the Struts framework) and a reference to the centrally maintained cache object, the LPS constructs a ProductServerRunner object (which is a sub-class of Java Thread). The Thread is started and the LPS joins this thread and waits for it complete. If it completes the request or encounters an error within the <ui_timeout> it returns the product or error to the client. If the Thread has not finished before the specified <ui_timeout> this Thread object is stored as an attribute in the Struts ActionServlet context and the LPS returns a progress page to the client (with information about how many times the progress page has refreshed since the request started and which step in the compound operation is currently being performed). After a 5 second delay, the browser client refreshes by sending the exact same request to the LPS. The LPS takes the UI Request and finds the Thread in its context that is processing this request. It joins that Thread and again waits for the <ui_timeout> to expire or the request to complete repeating the cycle of sending the progress page and refreshing the browser with the same request. Eventually, the ProductServerRunner will complete or the Backend Service will encounter the <product_timeout> and return a timeout error.

In the event that no <ui_timeout> is set in the incoming request (such as with a "batch" request from a command line client) the LPS will wait forever for the ProdcutServerRunner thread to complete without returning a progress page.

How the LPS Uses the LAS Configuration

The complexity of the operations and the mechanics of the data access necessary to make a product are hidden from the LUI by abstractions contained in XML configuration documents. For example, the configuration will contain details such as the units, time, geographic extents a particular data variable. The fact that these data reside on a remote OPeNDAP server or in a database table is hidden from the LUI. Likewise, the fact that creating the particular graphic product request might require the LPS to request that information be extracted from a database table and an intermediate netCDF file be created and then request a plot be made from this file is also hidden from the LUI.

Therefore, it is a large part of the job of the LPS to unwind these hidden complexities into the specific steps needed to create the product. It is not however, the job of the LPS to create the intermediate and final products needed to fulfill the request. Instead, the LPS will call upon one or more Backend Services to build the products necessary to complete a request. We have implemented a collection of Backend Services which we deploy in the LAS Backend Server (LBES). The LBES is nothing more than the Apache AxisServlet with a collection of LAS specific services deployed to the AxisServlet via the Axis WSDL.

In order to make the translation from the abstract references to data and operations included in the UI request to specific data access mechanisms and backend services, the LPS translates the UI request into an internal data structure called a ProductRequest. This translation highlights one of the first and most important design philosophies of the LPS architecture, the use of Plain Ordinary Java Objects (POJOs) to hold the application specific data.

The ProductRequest POJO[2] (a marriage between the UI Request and the LAS Configuration)

The constructor for the ProductRequest object takes as input the JDOM Document representing the LAS Configuration information and the JDOM Document representing UI Request XML. (Actually, all of the XML in the LPS and LBES are internally represented by sub-classes of LASDocument which is in turn a sub-class of JDOM Docment.) The ProductRequest object is independent of the Web application framework that is used to deploy the application. This means the JUnit testing can be used to exercise all of the code that creates and manipulates the ProductReqeust object and most importantly the underlying Web application framework can be changed without affecting a single line of the ProductReqeust object's code. The biggest collection of information in the ProductRequest object is an ArrayList<LASBackendRequest> which contains the XML documents that will used to invoke each backend service (some preliminary information about the LASBackendRequest XML document can be found here: http://tmap.pmel.noaa.gov/~rhs/JPS/LBES.html). When this XML is constructed all of the abstractions which point to data definitions in the LAS configuration are resolved into actual data access information either in form of file names or network resources (such as OPeNDAP URLs). Additionally, all of the property definitions which are contained in the LAS configuration XML are included in the XML that defines the data access for each LAS variable include in the request. Finally, this XML also contains concrete file names which will be the containers for the results this operation is expected to produce (plot images, map scale information, debug output and so on). The construction of the ProductRequest object is also responsible for creating the connections between compound operations where the results from one operation are used as input into subsequent operations.
The ProductActionRunner (the POJO that invokes the backend services)

When the ProductRequest object has been constructed, it is in turn used to construct the ProductServerRunner, the Thread sub-class that will be run to create the product. The ProductServerRunner Thread runs and sequentially invokes the services necessary to create the requested product and collates the responses.
Cache Mangement

All of the cache management occurs at the LPS before any backend services are invoked. The first check for a cache hit is based on the cached key generated from the UI Request. If there is a response file in the cache that matches this key, that response file is examined to make sure all of the individual results are still in the cache. If so, then the LPS returns this response immediately. If not it begins the process of invoking the individual services needed to produce the final product. Each individual sub-operation and its response are also cached so based on cache keys generated from the Backend Request XML for each sub-operation.

Individual backend service implementations are also welcome to manage a cache independently of the LPS, but are not required to do so.

The Backend Services (and how they are invoked and controlled)

The second important design philosophy has to do with the fact that the actions of the Backend Services that ultimately fulfill the product requests can be controlled by the LAS configuration. For the two services that we have implemented so far, this behavior is achieved by assigning a script template to each LAS operation.

For example, in the implementation of the Ferret backend service an operation that will perform a plot in a 2D plane is defined with this XML:

<operation name="2D Plot" ID="Plot_2D" template="output" script="Plot_2D">
<service>ferret</service>
<args>

<!-- This can now be improved with min/max info

to be used by the UI to enforce selections. -->
<arg type="variable" index="1" />
<arg type="region" index="1" />
</args>
<response ID="PlotResp" type="HTML" index="1">
<result type="image" ID="image1" />
<result type="image" ID="ref_map" />
<result type="map_scale" ID="map_scale" />
<result type="debug" ID="debug" />
<result type="cancel" ID="cancel" />
</response>
</operation>

The ProductServerRunner takes the ProductRequest and begins invoking the service for each LASBackendRequest. The SOAP call to the service is invoked and the LASBackendReqeust is passed to the backend service as a single String argument. The "script" attribute of this operation definition is a reference to a particular text file in the resource path of the backend service. When the backend service (named in the <service> element of the operation definition) receives this request, it will process this template as part of fulfilling the request (see below).

The Backend Services

At this point, we are ready to hop across the network to the backend services that are needed to create the product. Each of the backend services that we have implemented has the same basic structure. The service is a simple class which exposes a single method to the SOAP clients and this method accepts the LASBackendRequest XML string as its only argument. This request object is used to construct a Tool[1] class (which extends a generic Tool class). The Tool class must implement a run method which processes the request. If the Tool needs to interface with a legacy application such as Ferret a generic Task class is available which contains all of the infrastructure needed to run the external Unix application (methods for constructing the Runtime class to interface with the operating system, methods to pick up the output streams (which is critical since process can block with buffer overflows if this is not done correctly) and a generic methods to search the output streams for error conditions). I have used this Task class to run Ferret and to run Perl scripts.

When the service runs the aforementioned "script" file is what controls the actions of the legacy application or database query. This file is first processed as a Velocity template with the LASBackendRequest object in its context. All of the LAS configuration properties (that apply to this request or to the data variables referenced by this request) and names of input and output data files are available from the LASBackendRequest object and can thereby be translated into Ferret symbols or placed into SQL requests to customize the action for this particular request.

The backend service processes the request (runs the ferret or queries the db based on the "script") and constructs and LASBackendResponse object that will be returned to the LPS. This is a very simple XML construct that contains the names and IDs of the output files produced. (If the backend service is a remote service it must return a network URL instead of a file path in the response XML).

Canceling a Request

The backend service must also be prepared to handle a request to cancel a running request. If the LPS sends exactly the same request again (except containing a single <cancel/> element below the root element) the backend service should stop processing the request. Since getting access to the thread processing the backend request in the AxisServlet is complicated, we have implemented this action using the file system to signal that the request has been canceled. When the cancel request arrives the backend service creates an empty "cancel" file (defined in the operation XML) and immediately returns an error response indicating that the job has been canceled (even though the thread running the backend service may or may not have stopped processing the request at that moment). This means that the LPS can halt the ProductServerRunner that is processing the job and signal the user that the job has been canceled. Meanwhile, the original thread that is processing the job on the backend should be checking periodically (after completing various steps along the way in the Tool and Task classes) for the existence of the "cancel" file. If it finds the file it can stop processing and it has no further responsibilities and does not need to return any information to the LPS since the request that caused the cancel file to be created has already taken care of signaling the LPS and user that the job has been canceled.

Error Processing

If the backend service encounters an error processing the request it constructs a special LASBackendResponse with the error information included and the LPS will detect the error and forward to the generic error response (shown in the diagram below).
Processing the Response

We are now ready to return to the LSP. The LASBackendResponse is shipped back the ProductServerRunner (as a string) and the ProductServerRunner invokes the next service in the chain or signals the ProductServerAction that it is finished. The final task of the ProductActionServer is to process the response it receives. In any case, it removes the ProductActionRunner object from the Servlet context (at that moment subsequent requests for the same product will either get their request fulfilled from cache or will cause the product to be regenerated). If the response contains an error the action is forwarded to the error template which displays the error to the user. Otherwise, if the cache is turned on the LASBackendResponse file is added to the cache and the action is forwarded to the output template.

This last bit of work is the final part of the design philosophy in the new architecture. Every operation can have a template assigned to it which tells the LPS how to display the results. The template has access to the LASBackendResponse object and from this object it can extract any or all of the results that are defined in the response (be they debug text, plot output, ASCII data files and so on). It also has access to the LASUIRequest that was used to create this product so that it can make decisions about what to display, fill in titles and other information based on the values of properties that are defined for this request.

Finally, to allow the LPS to cause the browser to invoke external applications (such as Excel with CSV files) or to allow LPS URLs to used as the src attribute in <img> HTML tags, the LPS operations XML allows for operations to be defined with a single streaming result. E.g.

<result type="image" ID="plot_image" stream="true" mime_type="image/jpg"/>

If the LPS processes a request for an operation that asks for the output to be streamed, it sets the mime type of the response, picks up the bytes of the output and streams them back to the client then signals the Struts action servlet that is finished by forwarding to a null action.

The User Interface Server

[1] The Tool class and the Task class were stolen directly from Anagram and the implementations of Anagram used by TMAP. These classes were designed and written by Joe Wielgosz, Richard Rogers, Yonghua Wie.

[2] A POJO is a Plain Old Java Object as opposed to an Enterprise JavaBean or some other elaborate and specific object type.