-
Notifications
You must be signed in to change notification settings - Fork 0
Continuous Queries
Query tasks have a special option called Continuous queries. Continuous query is convienient way to run a long running query that at first sends initial results, and later additionally sends updates as the index get updated.
Developers may want to perform some form of aggregation based on the results of a query. In-order to keep the results up to date, the default approach will be periodically re-run the query and based on the results, re-run aggregation. This approach, even though works but is both resource intensive as well as incurs a delay depending on the interval the queries are run. Continuous query addresses this use case by providing a mechanism to developers to create a query task with a user defined query filter.
Continuous query is started by creating a query task service with CONTINUOUS option, and setting up a notification on that query task instance. A continuous query task does two things:
- It will self patch, causing the initial notification, once the initial query results are available, providing the documents currently in the index.
- On any additional index updates that match the query associated with the task, it will once again self patch, with the result being the document that changed or was added. Once again this will cause notifications to be sent.
In the remainder of the document we expect you have basic understanding of QueryTaskService.
The CONTINUOUS option creates a long running query filter that process all updates to the local index. The query specification is compiled into an efficient query filter that evaluates the document updates, and if the filter evaluates to true, the query task service is PATCHed with a document results reflecting the self-link (and document if EXPAND is set) that changed.
The continuous query task service acts as a node wide black board, or notification service allowing clients or services to receive notifications without having to subscribe to potentially millions of discrete services.
Here are the basic steps required to efficiently use the continuous query tasks.
- Create continuous query task request
- Send query task request
- On completion of the request, subscribe to the created query task service self-link
- Implement handler that will be called on notifications from query task service with updates
We can avoid setting up the subscription with query task here, and instead do the polling on this continues query task service for updates. But that would not be efficient. Instead we recommend using subscription model here and get the results whenever they are available.
In rest of the document we will go over the steps mentioned above.
Following is the simple example of QueryTask in which CONTINUOUS query option is selected. This query task is filtering the results based on Kind field clause.
public QueryTask createContinuousQuery() {
QueryTask.Query query = QueryTask.Query.Builder.create()
.addKindFieldClause(EmployeeService.Employee.class)
.build();
QueryTask queryTask = QueryTask.Builder.create()
.addOption(QueryOption.EXPAND_CONTENT)
.addOption(QueryOption.CONTINUOUS)
.setQuery(query).build();
return queryTask;
}By default, query tasks services expire after few (10 at the time of writing) minutes. If your continuous result collection is supposed to complete within that time limit then you are fine, otherwise you SHOULD increase this expiry limit to require time limit. Following code line is setting the expiry to be unlimited for query task service we are creating.
querytask.documentExpirationTimeMicros = Long.MAX_VALUE;
JSON payload of above query.
{
"taskInfo": {
"isDirect": false
},
"querySpec": {
"query": {
"occurance": "MUST_OCCUR",
"booleanClauses": [
{
"occurance": "MUST_OCCUR",
"term": {
"propertyName": "documentKind",
"matchValue": "com:vmware:myproject:EmployeeService:Employee",
"matchType": "TERM"
}
}
]
},
"options": [
"CONTINUOUS",
"EXPAND_CONTENT"
]
},
"indexLink": "/core/document-index",
"nodeSelectorLink": "/core/node-selectors/default",
"documentVersion": 0,
"documentUpdateTimeMicros": 0,
"documentExpirationTimeMicros": 0
}After creating the query task, we POST it to the local-query-tasks factory service with specific documentSelfLink (QUERY_TASK_LINK) that will create a query task with that name, helping us to point to this query task service easily in future.
On creation of this continuous query task service we call subscribeToContinuousQuery
to create the subscription to get notifications from query task service with updates.
Note: We always use a local query task (CORE_LOCAL_QUERY_TASKS) for continuous queries!
See also DeferredResult-Tutorial
public void createAndSubscribeToContinuousQuery(Operation op) {
QueryTask queryTask = createContinuousQuery();
queryTask.documentSelfLink = QUERY_TASK_LINK;
Operation post = Operation.createPost(getHost(), ServiceUriPaths.CORE_LOCAL_QUERY_TASKS)
.setBody(queryTask)
.setReferer(getHost().getUri());
getHost().sendWithDeferredResult(post)
.thenAccept((state) -> subscribeToContinuousQuery())
.whenCompleteNotify(op);
}Now let us see the code for subscribeToContinuousQuery which will create start
subscription service to listen for any notifications from continuous query task service
we created earlier. We are using startSubscriptionService API of the ServiceHost
to setup the subscription service. This API is creating a subscription service with a
callback URI, and registring that URI with the continuous query task service.
Please read the section "Important Details!" for mode details on creating a robust solution.
public void subscribeToContinuousQuery() {
Operation post = Operation
.createPost(UriUtils.buildUri(getHost(), QUERY_TASK_LINK))
.setReferer(getHost().getUri());
URI subscriptionUri = getHost().startSubscriptionService(post, this::processResults);
updateSubscriptionLink(subscriptionUri);
}See also section on Subscriptions in Programming-Model
Following is the basic implementation of processResults that would be called
whenever the query task service on the target host has any new data for us to
process. We check for presence of the results and then loop over all the result
documents to process. During the processing we can check state.documentUpdateAction
to see what was the last action (created, deleted) on that service.
public void processResults(Operation op) {
QueryTask body = op.getBody(QueryTask.class);
if (body.results == null || body.results.documentLinks.isEmpty()) {
return;
}
for (Object doc : body.results.documents.values()) {
EmployeeService.Employee state = Utils.fromJson(doc, EmployeeService.Employee.class);
getHost().log(Level.INFO, "Employee Name: %s, Action: %s", state.name, state.documentUpdateAction);
}
}So far we have went over the basics of continuous query tasks. But there are few things, we need to be aware of, when we are writing code for production.
In a multi-node environment, one important thing to note is that continuous query task service, being a long running service, that hits the index regularly for updates. Hence this should be used with care and should be called by a single host. If above code is running on a replicated service on multiple nodes, then we would be triggering this continuous query task multiple times on our host, which would be a over kill and can be big performance hit on your hosts.
That means when we are creating continuous queries, we need to make sure, that the caller with subscription to notifications is fault-tolarant and load-balanced. We do that by making a stateful service which is replicated, and owner-selected. Lets call this service a watch service.
Because it will be a replicated service, it will give us fault-tolerance in case one of the replica goes down. And beause this is owner-selected we can make sure that only one node is doing the heavy lifting at anyone time, and other nodes are not dublicating the work. That watch service will create the continuous query task and watch for notifications from it. But not just that, we would want it to do this only on one of the replicated nodes, not on all nodes. How do we make sure that stateful service is only creating the query task in one of the replcas? We can do that by checking for the ownership of its state in handleNodeGroupMaintenance method.
What is handleNodeGroupMaintenance? This is overridable method in stateful
services that is invoked by the Xenon runtime when there is change in the
ownership of the state of the service.
We can override this method in our watch service and check if we are owner or
not using hasOption(ServiceOption.DOCUMENT_OWNER), and based on that create
continuous query task if we are owner, and delete running continuous query task
service and subscription if we are not the owner anymore.
Following is the example code that implements handleNodeGroupMaintenance
@Override
public void handleNodeGroupMaintenance(Operation op) {
// Create continuous queries and subscriptions in case of change in node group topology.
if (hasOption(ServiceOption.DOCUMENT_OWNER)) {
createAndSubscribeToContinuousQuery(op);
} else {
deleteSubscriptionAndContinuousQuery(op);
}
}
We have already seen the basic implementation of
createAndSubscribeToContinuousQueryTask earlier. Now lets see what
deleteSubscriptionAndContinuousQuery might look like in this watch service.
private void deleteSubscriptionAndContinuousQuery(Operation op) {
Operation unsubscribeOperation = Operation.createPost(UriUtils.buildUri(getHost(), QUERY_TASK_LINK))
.setReferer(getUri())
.setCompletion((o, e) -> {
updateSubscriptionLink(null);
deleteContinuousQuery();
});
getStateAndApply(state -> getHost().stopSubscriptionService(unsubscribeOperation,
UriUtils.buildUri(state.subscriptionLink)));
}
In above method we are first creating a Operation to unsubscribe the
subscription service being created by the runtime. stopSubscriptionService deletes
both the subscriber entry in the query task service and also deletes the subscription callback
endpoint.
On the completion of these deletions we are deleting the continuous query service by
calling deleteContinuousQuery.
Following is the implementation of deleteContinuousQuery.
private void deleteContinuousQuery() {
getHost().sendRequest(Operation
.createDelete(UriUtils.buildUri(getHost(), QUERY_TASK_LINK))
.setReferer(getUri()));
}
Following is the implementation of getStateAndApply.
private void getStateAndApply(Consumer<? super State> action) {
Operation get = Operation
.createGet(this, this.getSelfLink())
.setReferer(getUri());
getHost().sendWithDeferredResult(get, State.class)
.thenAccept(action)
.whenCompleteNotify(get);
}
No, because queries work on stateful and persisted services.
You can do curl on http://[host]/core/local-query-tasks to see list of all local query task services.
It got expired after 10 minutes. Set it to never expire using querytask.documentExpirationTimeMicros = Long.MAX_VALUE;
Subscription handler will be called when there are any updates. Make sure your updates are being reflected on the index.
Let’s say I’ve got a query for “all example services”, and one of the example services is deleted: do I learn that it’s no longer in the query result?
You will get a PATCH, with the body being the last version of the example
service when it was deleted. The documentUpdateAction will be DELETE.
You said, any update that satisfies the query filter will cause the results to be updated and a self PATCH to be sent on the service. Does that mean I’ll see a diff of new things that match that query?
You will receive notifications in the form of PATCH operations, with the body being a QueryTask, with the results.documentLinks/documents being filled in with the specific update to a service that matched the query if your query no longer match anything, you get no notifications. You can cancel it, have it expire, etc.
It never ends, until it gets expired. You will keep getting notification as long as there are updates that fulfil continuous query task's filter or until it gets expired.
How do you calculate a total sum using a continuous query when you don't know when your are done getting update notifications?
Well, a total sum implies you know the full set that you want to compute it over, which means that you can use a normal query. If you want to keep counting, and do a running sum, then you need to use a continuous query.