io.mats3.util.MatsFuturizer Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of mats-util Show documentation
Show all versions of mats-util Show documentation
Mats^3 Utilities - notably the MatsFuturizer, which provides a bridge from synchronous processes to the highly asynchronous Mats^3 services.
package io.mats3.util;
import java.util.ArrayList;
import java.util.HashMap;
import java.util.List;
import java.util.PriorityQueue;
import java.util.concurrent.CompletableFuture;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.LinkedTransferQueue;
import java.util.concurrent.ThreadPoolExecutor;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.TimeoutException;
import java.util.concurrent.atomic.AtomicInteger;
import java.util.concurrent.locks.Condition;
import java.util.concurrent.locks.ReentrantLock;
import java.util.function.Function;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import org.slf4j.MDC;
import io.mats3.MatsEndpoint;
import io.mats3.MatsEndpoint.DetachedProcessContext;
import io.mats3.MatsEndpoint.MatsObject;
import io.mats3.MatsEndpoint.ProcessContext;
import io.mats3.MatsEndpoint.ProcessTerminatorLambda;
import io.mats3.MatsFactory;
import io.mats3.MatsFactory.FactoryConfig;
import io.mats3.MatsInitiator;
import io.mats3.MatsInitiator.InitiateLambda;
import io.mats3.MatsInitiator.MatsInitiate;
/**
* An instance of this class acts as a bridge service between the synchronous world of e.g. a HTTP request, and the
* asynchronous world of Mats. In a given project, you typically create a singleton instance of this class upon startup,
* and employ it for all such scenarios. In short, in a HTTP service handler, you initialize a Mats flow using
* {@link #futurizeNonessential(CharSequence, String, String, Class, Object)
* singletonFuturizer.futurizeNonessential(...)} (or
* {@link #futurize(CharSequence, String, String, int, TimeUnit, Class, Object, InitiateLambda) futurize(...)} for full
* configurability), specifying which Mats Endpoint to invoke and the request DTO instance, and then you get a
* {@link CompletableFuture} in return. This future will complete once the invoked Mats Endpoint replies.
*
* It is extremely important to understand that this is NOT how you compose multiple Mats Endpoints together! This is
* ONLY supposed to be used when you are in a synchronous context (e.g. in a Servlet, or a Spring @RequestMapping) "on
* the edge" of the Mats fabric, and want to interact with the Mats fabric of Endpoints.
*
* Another aspect to understand, is that while Mats "guarantees" that a successfully submitted initiation will flow
* through the Mats endpoints, no matter what happens with the processing nodes (unless you employ NonPersistent
* messaging, which futurizeNonessential(..) does!), nothing can be guaranteed wrt. the completion of the
* future: This is stateful processing. The node where the MatsFuturizer initiation is performed can crash right after
* the message has been put on the Mats fabric, and hence the CompletableFuture vanishes along with everything else on
* that node. The mats flow is however already in motion, and will be executed - but when the Reply comes in on the
* node-specific Topic, there is no longer any corresponding CompletableFuture to complete. This is also why you should
* not compose Mats endpoints using this familiar feeling that a CompletableFuture probably gives you: While a
* multi-stage MatsEndpoint is asynchronous, resilient and highly available and each stage is transactionally performed,
* with retries and all the goodness that comes with a message oriented architecture, once you rely on a
* CompletableFuture, you are in a synchronous world where a power outage or a reboot can stop the processing midway.
* Thus, the MatsFuturizer should always just be employed out the very outer edge facing the actual client - any other
* processing should be performed using MatsEndpoints, and composition of MatsEndpoints should be done using multi-stage
* MatsEndpoints.
*
* Note that in the case of pure "GET-style" requests where information is only retrieved and no state in the total
* system is changed, everything is a bit more relaxed: If a processing fails, the worst thing that happens is a
* slightly annoyed user. But if this was an "add order" or "move money" instruction from the user, a mid-processing
* failure is rather bad and could require human intervention to clean up. Thus, the
* futurizeNonessential(..) method should only be employed for such safe "GET-style" requests.
* Any other potentially state changing operations must employ the generic futurize(..) method.
*
* A question you might have, is how this works in a multi-node setup? For a Mats flow, it does not matter which node a
* given stage of a MatsEndpoint is performed, as it is by design totally stateless wrt. the executing node, as all
* state resides in the message. However, for a synchronous situation as in a HTTP request, it definitely matters that
* the final reply, the one that should complete the returned future, comes in on the same node that issued the request,
* as this is where the CompletableFuture instance is, and where the waiting TCP connection is connected! The trick here
* is that the final reply is specified to come in on a node-specific topic, i.e. it literally has the node name
* (default being the hostname) as a part of the MatsEndpoint name, and it is a
* {@link MatsFactory#subscriptionTerminator(String, Class, Class, ProcessTerminatorLambda) SubscriptionTerminator}.
*
* Logger MDCs for completion and metrics (on the logger "io.mats3.util.MatsFuturizer.Reply" if
* INFO-enabled):
*
* - {@link #MDC_MATS_FUTURE_COMPLETED "mats.FutureCompleted"}: Present on a single logline per Future
* completed, the value is the total time taken from futurization Request was initiated, until the future is
* completed.
* - {@link #MDC_TRACE_ID "traceId"}: The TraceId the futurization was initiated with.
* - {@link #MDC_MATS_INIT_ID "mats.init.Id"}: The 'from' parameter in the futurization call, i.e. the
* initiatorId
* - {@link #MDC_MATS_FUTURE_TIME_RTT "mats.future.rtt.ms"}: Part of the total time used for the Mats3 round
* trip from futurization Request was initiated, through the internal SubscriptionTerminator received the Reply, until
* the Futurizer's thread pool created the Reply-instance.
* - {@link #MDC_MATS_FUTURE_TIME_COMPLETING "mats.future.completing.ms"}: Part of the total time used to
* complete the future. If the calling thread that initiated the futurization directly blocks on the future.get(), this
* value will be very close to zero. However, if there are thenApplys and/or thenAccepts involved, those will increase
* this time.
*
* Logger for MDCs for timeouts:
*
* - {@link #MDC_MATS_FUTURE_TIMEOUT "mats.FutureTimeout"}: Present on a single logline when a Future is timed
* out by the MatsFuturizer, for oversitting its specified timeout upon futurization initiation. The value is the time
* since it was initiated.
* - {@link #MDC_TRACE_ID "traceId"}: Same as completed.
* - {@link #MDC_MATS_INIT_ID "mats.init.Id"}: Same as completed.
*
*
* @author Endre Stølsvik 2019-08-25 20:35 - http://stolsvik.com/, [email protected]
*/
public class MatsFuturizer implements AutoCloseable {
private static final Logger log = LoggerFactory.getLogger(MatsFuturizer.class);
private static final String LOG_PREFIX = "#MATS-UTIL# ";
public static final String MDC_TRACE_ID = "traceId";
public static final String MDC_MATS_INIT_ID = "mats.init.Id"; // matsInitiate.from(initiatorId).
public static final String MDC_MATS_FUTURE_COMPLETED = "mats.FutureCompleted";
public static final String MDC_MATS_FUTURE_TIME_RTT = "mats.future.rtt.ms";
public static final String MDC_MATS_FUTURE_TIME_COMPLETING = "mats.future.completing.ms";
public static final String MDC_MATS_FUTURE_TIMEOUT = "mats.FutureTimeout";
/**
* Creates a MatsFuturizer, and you should only need one per MatsFactory (which again mostly means one per
* application or micro-service or JVM). The defaults for the parameters from the fully fledged factory method are
* identical to the {@link #createMatsFuturizer(MatsFactory, String)}, but with this variant also the
* 'endpointIdPrefix' is set to what is returned by matsFactory.getFactoryConfig().getAppName().
* Note that if you - against the above suggestion - create more than one MatsFuturizer for a MatsFactory, then
* you MUST give them different endpointIdPrefixes, thus you cannot use this method!
*
* @param matsFactory
* the underlying {@link MatsFactory} on which outgoing messages will be sent, and on which the receiving
* {@link MatsFactory#subscriptionTerminator(String, Class, Class, ProcessTerminatorLambda)
* SubscriptionTerminator} will be created.
* @return the {@link MatsFuturizer}, which is tied to a newly created
* {@link MatsFactory#subscriptionTerminator(String, Class, Class, ProcessTerminatorLambda)
* SubscriptionTerminator}.
*/
public static MatsFuturizer createMatsFuturizer(MatsFactory matsFactory) {
String endpointIdPrefix = matsFactory.getFactoryConfig().getAppName();
if ((endpointIdPrefix == null) || endpointIdPrefix.trim().isEmpty()) {
throw new IllegalArgumentException("The matsFactory.getFactoryConfig().getAppName() returns ["
+ endpointIdPrefix + "], which is not allowed to use as endpointIdPrefix (null or blank).");
}
return createMatsFuturizer(matsFactory, endpointIdPrefix);
}
/**
* Creates a MatsFuturizer, and you should only need one per MatsFactory (which again mostly means one per
* application or micro-service or JVM). The number of threads in the future-completer-pool is what
* {@link FactoryConfig#getConcurrency() matsFactory.getFactoryConfig().getConcurrency()} returns at creation time x
* 4 for "corePoolSize", but at least 5, (i.e. "min"); and concurrency * 20, but at least 100, for "maximumPoolSize"
* (i.e. max). The pool is set up to let non-core threads expire after 5 minutes. The maximum number of outstanding
* promises is set to 50k.
*
* @param matsFactory
* the underlying {@link MatsFactory} on which outgoing messages will be sent, and on which the receiving
* {@link MatsFactory#subscriptionTerminator(String, Class, Class, ProcessTerminatorLambda)
* SubscriptionTerminator} will be created.
* @param endpointIdPrefix
* the first part of the endpointId, which typically should be some "class-like" construct denoting the
* service name, like "OrderService" or "InventoryService", preferably the same prefix you use for all
* your other endpoints running on this same service. Note: If you create multiple MatsFuturizers for
* a MatsFactory, this parameter must be different for each instance!
* @return the {@link MatsFuturizer}, which is tied to a newly created
* {@link MatsFactory#subscriptionTerminator(String, Class, Class, ProcessTerminatorLambda)
* SubscriptionTerminator}.
*/
public static MatsFuturizer createMatsFuturizer(MatsFactory matsFactory, String endpointIdPrefix) {
int corePoolSize = Math.max(5, matsFactory.getFactoryConfig().getConcurrency() * 4);
int maximumPoolSize = Math.max(100, matsFactory.getFactoryConfig().getConcurrency() * 20);
return createMatsFuturizer(matsFactory, endpointIdPrefix, corePoolSize, maximumPoolSize, 50_000);
}
/**
* Creates a MatsFuturizer, and you should only need one per MatsFactory (which again mostly means one per
* application or micro-service or JVM). With this factory method you can specify the number of threads in the
* future-completer-pool with the parameters "corePoolSize" and "maxPoolSize" threads, which effectively means min
* and max. The pool is set up to let non-core threads expire after 5 minutes. You must also specify the max number
* of outstanding promises, if you want no effective limit, use {@link Integer#MAX_VALUE}.
*
* @param matsFactory
* the underlying {@link MatsFactory} on which outgoing messages will be sent, and on which the receiving
* {@link MatsFactory#subscriptionTerminator(String, Class, Class, ProcessTerminatorLambda)
* SubscriptionTerminator} will be created.
* @param endpointIdPrefix
* the first part of the endpointId, which typically should be some "class-like" construct denoting the
* service name, like "OrderService" or "InventoryService", preferably the same prefix you use for all
* your other endpoints running on this same service. Note: If you create multiple MatsFuturizers for
* a MatsFactory, this parameter must be different for each instance!
* @param corePoolSize
* the minimum number of threads in the future-completer-pool of threads.
* @param maxPoolSize
* the maximum number of threads in the future-completer-pool of threads.
* @param maxOutstandingPromises
* the maximum number of outstanding Promises before new are rejected. Should be a fairly high number,
* e.g. the default of {@link #createMatsFuturizer(MatsFactory, String)} is 50k.
* @return the {@link MatsFuturizer}, which is tied to a newly created
* {@link MatsFactory#subscriptionTerminator(String, Class, Class, ProcessTerminatorLambda)
* SubscriptionTerminator}.
*/
public static MatsFuturizer createMatsFuturizer(MatsFactory matsFactory, String endpointIdPrefix,
int corePoolSize, int maxPoolSize, int maxOutstandingPromises) {
return new MatsFuturizer(matsFactory, endpointIdPrefix, corePoolSize, maxPoolSize, maxOutstandingPromises);
}
protected final MatsFactory _matsFactory;
protected final MatsInitiator _matsInitiator;
protected final String _terminatorEndpointId;
protected final ThreadPoolExecutor _futureCompleterThreadPool;
protected final int _maxOutstandingPromises;
protected final MatsEndpoint _replyHandlerEndpoint;
protected MatsFuturizer(MatsFactory matsFactory, String endpointIdPrefix, int corePoolSize, int maxPoolSize,
int maxOutstandingPromises) {
_matsFactory = matsFactory;
String endpointIdPrefix_sanitized = SanitizeMqNames.sanitizeName(endpointIdPrefix);
if ((endpointIdPrefix_sanitized == null) || endpointIdPrefix_sanitized.trim().isEmpty()) {
throw new IllegalArgumentException("The sanitized endpointIdPrefix (orig:["
+ endpointIdPrefix + "]) is not allowed to use as endpointIdPrefix (null or blank).");
}
_matsInitiator = matsFactory.getOrCreateInitiator(endpointIdPrefix_sanitized + ".Futurizer.init");
_terminatorEndpointId = endpointIdPrefix_sanitized + ".Futurizer.private.repliesFor."
+ _matsFactory.getFactoryConfig().getNodename();
_futureCompleterThreadPool = _newThreadPool(corePoolSize, maxPoolSize);
_maxOutstandingPromises = maxOutstandingPromises;
_replyHandlerEndpoint = _matsFactory.subscriptionTerminator(_terminatorEndpointId, String.class,
MatsObject.class,
this::_handleRepliesForPromises);
_startTimeouterThread();
log.info(LOG_PREFIX + "MatsFuturizer created."
+ " EndpointIdPrefix:[" + endpointIdPrefix_sanitized
+ "], corePoolSize:[" + corePoolSize
+ "], maxPoolSize:[" + maxPoolSize
+ "], maxOutstandingPromises:[" + maxOutstandingPromises + "]");
}
/**
* An instance of this class will be the return value of the {@link CompletableFuture}s created by the
* {@link MatsFuturizer}. It will contain the reply from the requested endpoint, and the
* {@link DetachedProcessContext DetachedProcessContext} from the received message, from where you can get any
* incoming {@link DetachedProcessContext#getBytes(String) "sideloads"} and other metadata. It also contains a
* timestamp of when the outgoing message was initiated (both as {@link #getInitiationTimestamp() millis} and
* {@link #getInitiationNanos() nanos}), as well as the {@link #getRoundTripNanos() round-trip-time} in nanos.
*
* You may choose between using the final fields, or the getters, as you prefer. You should probably be consistent
* within a project!
*
* @param
* the type of the reply class.
*/
public static class Reply {
private static final Logger log = LoggerFactory.getLogger(Reply.class);
/**
* The {@link DetachedProcessContext} from the received message, from where you can get any incoming
* {@link DetachedProcessContext#getBytes(String) "sideloads"} and other metadata.
*/
public final DetachedProcessContext context;
/**
* The actual Reply DTO from the requested Endpoint
*/
public final T reply;
/**
* When this request was initiated, from {@link System#currentTimeMillis() System.currentTimeMillis()}.
*/
public final long initiationTimestamp;
/**
* When this request was initiated, from {@link System#nanoTime() System.nanoTime()}.
*/
public final long initiationNanos;
/**
* The number of nanos between the Request was sent to targeted Endpoint, and when MatsFuturizer received the
* Reply from the Endpoint on the internal SubscriptionTerminator.
*/
public final long roundTripNanos;
/**
* SOFT DEPRECATED, constructor available for legacy reasons. Please move away! Use the new forTest(..) factory
* methods instead.
*/
public Reply(DetachedProcessContext context, T reply, long initiationTimestamp) {
log.warn(LOG_PREFIX + "HARD WARNING - DEPRECATION!! Using the new Reply(context, reply,"
+ " initiationTimestamp) constructor is deprecated, use Reply.forTest(context, reply) instead!");
this.context = context;
this.reply = reply;
this.initiationTimestamp = initiationTimestamp;
this.initiationNanos = 0;
this.roundTripNanos = 0;
}
/**
* Factory method for testing scenarios, where you want to create a Reply instance only containing the reply.
* Timestamps and RTT will be zero, and the DetachedProcessContext will be null.
*/
public static Reply forTest(T reply) {
return new Reply<>(null, reply, 0, 0);
}
/**
* Factory method for testing scenarios, where you want to create a Reply instance only containing the reply and
* the DetachedProcessContext. If you need anything more, you should use Mockito. Timestamps and RTT will be
* zero.
*/
public static Reply forTest(DetachedProcessContext context, T reply) {
return new Reply<>(context, reply, 0, 0);
}
private Reply(DetachedProcessContext context, T reply, long initiationTimestamp, long initiationNanos) {
this.context = context;
this.reply = reply;
this.initiationTimestamp = initiationTimestamp;
this.initiationNanos = initiationNanos;
this.roundTripNanos = System.nanoTime() - initiationNanos;
}
/**
* @return the {@link DetachedProcessContext} from the received message, from where you can get any incoming
* {@link DetachedProcessContext#getBytes(String) "sideloads"} and other metadata.
*/
public DetachedProcessContext getContext() {
return context;
}
/**
* SOFT DEPRECATED, use {@link #get()}.
*/
public T getReply() {
log.warn(LOG_PREFIX + "HARD WARNING - DEPRECATION!! Using Reply.getReply() is deprecated,"
+ " use Reply.get()!");
return get();
}
/**
* @return the actual Reply DTO from the requested Endpoint
*/
public T get() {
return reply;
}
/**
* @return when this request was initiated, from {@link System#currentTimeMillis() System.currentTimeMillis()}.
*/
public long getInitiationTimestamp() {
return initiationTimestamp;
}
/**
* @return when this request was initiated, from {@link System#nanoTime() System.nanoTime()}.
*/
public long getInitiationNanos() {
return initiationNanos;
}
/**
* @return the number of nanos between the Request was sent to targeted Endpoint, and when MatsFuturizer
* received the Reply from the Endpoint on the internal SubscriptionTerminator.
*/
public long getRoundTripNanos() {
return roundTripNanos;
}
}
/**
* This exception is raised through the {@link CompletableFuture} if the timeout specified when getting the
* {@link CompletableFuture} is reached (to get yourself a future, use one of the
* {@link #futurize(CharSequence, String, String, Class, Object, InitiateLambda) futurize(..)} methods). The
* exception is passed to the waiter on the future by {@link CompletableFuture#completeExceptionally(Throwable)},
* where the consumer can pick it up with e.g. {@link CompletableFuture#exceptionally(Function)}.
*/
public static class MatsFuturizerTimeoutException extends RuntimeException {
private final long initiationTimestamp;
private final String traceId;
public MatsFuturizerTimeoutException(String message, long initiationTimestamp, String traceId) {
super(message);
this.initiationTimestamp = initiationTimestamp;
this.traceId = traceId;
}
public long getInitiationTimestamp() {
return initiationTimestamp;
}
public String getTraceId() {
return traceId;
}
}
/**
* The generic form of initiating a request-message that returns a {@link CompletableFuture}, which enables you to
* tailor all properties. To set interactive-, nonPersistent- or noAudit-flags, or to tack on any
* {@link MatsInitiate#addBytes(String, byte[]) "sideloads"} to the outgoing message, use the "customInit"
* parameter, which directly is the {@link InitiateLambda InitiateLambda} that the MatsFuturizer initiation is
* using.
*
* For a bit more explanation, please read JavaDoc of
* {@link #futurizeNonessential(CharSequence, String, String, Class, Object) futurizeInteractiveUnreliable(..)}
*
* @param traceId
* TraceId of the resulting Mats call flow, see {@link MatsInitiate#traceId(CharSequence)}
* @param from
* the "from" of the initiation, see {@link MatsInitiate#from(String)}
* @param to
* to which Mats endpoint the request should go, see {@link MatsInitiate#to(String)}
* @param timeout
* how long before the internal timeout-mechanism of MatsFuturizer kicks in and the future is
* {@link CompletableFuture#completeExceptionally(Throwable) completed exceptionally} with a
* {@link MatsFuturizerTimeoutException}.
* @param unit
* the unit of time of the 'timeout' parameter.
* @param replyClass
* which expected reply DTO class that the requested endpoint replies with.
* @param request
* the request DTO that should be sent to the endpoint, see {@link MatsInitiate#request(Object)}
* @param customInit
* the {@link InitiateLambda} that the MatsFuturizer is employing to initiate the outgoing message, which
* you can use to tailor the message, e.g. setting the {@link MatsInitiate#interactive()
* interactive}-flag or tacking on {@link MatsInitiate#addBytes(String, byte[]) "sideloads"}.
* @param
* the type of the reply DTO.
* @return a {@link CompletableFuture} which will be resolved with a {@link Reply}-instance that contains both some
* meta-data, and the {@link Reply#get() reply} from the requested endpoint.
*/
public CompletableFuture> futurize(CharSequence traceId, String from, String to,
int timeout, TimeUnit unit, Class replyClass, Object request, InitiateLambda customInit) {
Promise promise = _createPromise(traceId.toString(), from, to, replyClass, timeout, unit);
_assertFuturizerRunning();
_enqueuePromise(promise);
_sendRequestToFulfillPromise(from, to, traceId.toString(), request, customInit, promise);
return promise._future;
}
/**
* Convenience-variant of the generic
* {@link #futurize(CharSequence, String, String, int, TimeUnit, Class, Object, InitiateLambda) futurize(..)} form,
* where the timeout is set to 2.5 minutes. To set interactive-, nonPersistent- or noAudit-flags, or to tack on any
* {@link MatsInitiate#addBytes(String, byte[]) "sideloads"} to the outgoing message, use the "customInit"
* parameter, which directly is the {@link InitiateLambda InitiateLambda} that the MatsFuturizer initiation is
* using.
*
* For a bit more explanation, please read JavaDoc of
* {@link #futurizeNonessential(CharSequence, String, String, Class, Object) futurizeInteractiveUnreliable(..)}
*
* @param traceId
* TraceId of the resulting Mats call flow, see {@link MatsInitiate#traceId(CharSequence)}
* @param from
* the "from" of the initiation, see {@link MatsInitiate#from(String)}
* @param to
* to which Mats endpoint the request should go, see {@link MatsInitiate#to(String)} the unit of time of
* the 'timeout' parameter.
* @param replyClass
* which expected reply DTO class that the requested endpoint replies with.
* @param request
* the request DTO that should be sent to the endpoint, see {@link MatsInitiate#request(Object)}
* @param customInit
* the {@link InitiateLambda} that the MatsFuturizer is employing to initiate the outgoing message, which
* you can use to tailor the message, e.g. setting the {@link MatsInitiate#interactive()
* interactive}-flag or tacking on {@link MatsInitiate#addBytes(String, byte[]) "sideloads"}.
* @param
* the type of the reply DTO.
* @return a {@link CompletableFuture} which will be resolved with a {@link Reply}-instance that contains both some
* meta-data, and the {@link Reply#get() reply} from the requested endpoint.
*/
public CompletableFuture> futurize(CharSequence traceId, String from, String to, Class replyClass,
Object request, InitiateLambda customInit) {
return futurize(traceId, from, to, 150, TimeUnit.SECONDS, replyClass, request, customInit);
}
/**
* NOTICE: This variant must only be used for "GET-style" Requests where none of the endpoints the call
* flow passes will add, remove or alter any state of the system, and where it doesn't matter all that much if a
* message (and hence the Mats flow) is lost!
*
* The goal of this method is to be able to get hold of e.g. account holdings, order statuses etc, for presentation
* to a user. The thinking is that if such a flow fails where a message of the call flow disappears, this won't make
* for anything else than a bit annoyed user: No important state change, like the adding, deleting or change of an
* order, will be lost. Also, speed is of the essence. Therefore, non-persistent. At the same time, to make
* the user super happy in the ordinary circumstances, all messages in this call flow will be prioritized, and thus
* skip any queue backlogs that have arose on any of the call flow's endpoints, e.g. due to some massive batch of
* (background) processes executing at the same time. Therefore, interactive. Notice that with both of these
* features combined, you get very fast messaging, as non-persistent means that the message will not have to be
* stored to permanent storage at any point, while interactive means that it will skip any backlogged queues. In
* addition, the noAudit flag is set, since it is a waste of storage space to archive the actual contents of
* Request and Reply messages that do not alter the system.
*
* Sets the following properties on the sent Mats message:
*
* - Non-persistent: Since it is not vitally important that this message is not lost, non-persistent
* messaging can be used. The minuscule chance for this message to disappear is not worth the considerable overhead
* of store-and-forward multiple times to persistent storage. Also, speed is much more interesting.
* - Interactive: Since the Futurizer should only be used as a "synchronous bridge" when a human is
* actively waiting for the response, the interactive flag is set. (For all other users, you should rather code
* "proper Mats" with initiations, endpoints and terminators).
* - No audit: Since this message will not change the state of the system (i.e. the "GET-style" requests),
* using storage on auditing requests and replies is not worthwhile.
*
* This method initiates an {@link MatsInitiate#nonPersistent() non-persistent} (unreliable),
* {@link MatsInitiate#interactive() interactive} (prioritized), {@link MatsInitiate#noAudit()
* non-audited} (request and reply DTOs won't be archived) Request-message to the specified endpoint, returning
* a {@link CompletableFuture} that will be {@link CompletableFuture#complete(Object) completed} when the Reply from
* the requested endpoint comes back. The internal MatsFuturizer timeout will be set to 2.5 minutes, meaning
* that if there is no reply forthcoming within that time, the {@link CompletableFuture} will be
* {@link CompletableFuture#completeExceptionally(Throwable) completed exceptionally} with a
* {@link MatsFuturizerTimeoutException MatsFuturizerTimeoutException}, and the Promise deleted from the futurizer.
* 2.5 minutes is probably too long to wait for any normal interaction with a system, so if you use the
* {@link CompletableFuture#get(long, TimeUnit) CompletableFuture.get(timeout, TimeUnit)} method of the returned
* future, you might want to put a lower timeout there - if the answer hasn't come within that time, you'll get a
* {@link TimeoutException}. If you instead use the non-param variant {@link CompletableFuture#get() get()}, you
* will get an {@link ExecutionException} when the 2.5 minutes have passed (that exception's
* {@link ExecutionException#getCause() cause} will be the {@link MatsFuturizerTimeoutException
* MatsFuturizerTimeoutException} mentioned above).
*
* @param traceId
* TraceId of the resulting Mats call flow, see {@link MatsInitiate#traceId(CharSequence)}
* @param from
* the "from" of the initiation, see {@link MatsInitiate#from(String)}
* @param to
* to which Mats endpoint the request should go, see {@link MatsInitiate#to(String)}
* @param replyClass
* which expected reply DTO class that the requested endpoint replies with.
* @param request
* the request DTO that should be sent to the endpoint, see {@link MatsInitiate#request(Object)}
* @param
* the type of the reply DTO.
* @return a {@link CompletableFuture} which will be resolved with a {@link Reply}-instance that contains both some
* meta-data, and the {@link Reply#get() reply} from the requested endpoint.
*/
public CompletableFuture> futurizeNonessential(CharSequence traceId, String from, String to,
Class replyClass, Object request) {
// Using 150 seconds (2.5 min) as default timeout, with 180 seconds (3 min) as TTL
return futurize(traceId, from, to, 150, TimeUnit.SECONDS, replyClass, request,
msg -> msg.nonPersistent(180_000).interactive().noAudit());
}
/**
* @return the number of outstanding promises, not yet completed or timed out.
*/
public int getOutstandingPromiseCount() {
_internalStateLock.lock();
try {
return _correlationIdToPromiseMap.size();
}
finally {
_internalStateLock.unlock();
}
}
/**
* @return the future-completer-thread-pool, for introspection. If you mess with it, you will be sorry..!
*/
public ThreadPoolExecutor getCompleterThreadPool() {
return _futureCompleterThreadPool;
}
// ===== Internal classes and methods, can be overridden if you want to make a customized MatsFuturizer
// .. but that is on your own risk - this is not a public API per se, and may change.
protected static class Promise implements Comparable> {
public final String _traceId;
public final String _correlationId;
public final String _from;
public final String _to;
public final long _initiationTimestamp;
public final long _initiationNanos;
public final long _timeoutTimestamp;
public final Class _replyClass;
public final CompletableFuture> _future;
public Promise(String traceId, String correlationId, String from, String to, long initiationTimestamp,
long initiationNanos, long timeoutTimestamp, Class replyClass, CompletableFuture> future) {
_traceId = traceId;
_correlationId = correlationId;
_from = from;
_to = to;
_initiationTimestamp = initiationTimestamp;
_initiationNanos = initiationNanos;
_timeoutTimestamp = timeoutTimestamp;
_replyClass = replyClass;
_future = future;
}
@Override
public int compareTo(Promise o) {
// ?: Are timestamps equal?
if (this._timeoutTimestamp == o._timeoutTimestamp) {
// -> Yes, timestamps equal, so compare by correlationId.
return this._correlationId.compareTo(o._correlationId);
}
// "signum", but zero is handled above.
return this._timeoutTimestamp - o._timeoutTimestamp > 0 ? +1 : -1;
}
}
protected final AtomicInteger _threadNumber = new AtomicInteger();
protected ThreadPoolExecutor _newThreadPool(int corePoolSize, int maximumPoolSize) {
// Trick to make ThreadPoolExecutor work as anyone in the world would expect:
// Have a constant pool of "corePoolSize", and then as more tasks are concurrently running than threads
// available, you increase the number of threads until "maximumPoolSize", at which point the rest go on queue.
// Snitched from https://stackoverflow.com/a/24493856, with a twist to make it work with Java 21.0.2.
/*
* Part 1: So, we extend a LinkedTransferQueue to behave a bit special on "offer(..)": The ThreadPoolExecutor
* (TPE) will when it has exhausted the core pool size, put the task on queue via "offer(E)". But, in offer(E),
* we'll try to "forcefully" give the task to the TPE, so that it starts scaling the pool towards the maximum.
* However, if the TPE is "full" (i.e. it has reached maximumPoolSize), it will return false, and the task will
* be rejected. We'll then have a RejectionExecutionHandler that puts the task on queue anyway, even if it's
* "full".
*
* Previous to Java 21.0.2, we used the "put(E)" method to put stuff on queue. However, with Java 21.0.2, the
* "put(E)" method was reimplemented to directly call "offer(E)", so we could not anymore rely on "put(E)" to
* put stuff on queue. Therefore, we'll make a special "sneaky" method to put stuff on queue.
*/
class LinkedTransferQueueSneaky extends LinkedTransferQueue {
@Override
public boolean offer(E e) {
return tryTransfer(e);
}
public void sneak(E e) {
super.offer(e);
}
}
LinkedTransferQueueSneaky queue = new LinkedTransferQueueSneaky<>();
ThreadPoolExecutor threadPool = new ThreadPoolExecutor(corePoolSize, maximumPoolSize,
5L, TimeUnit.MINUTES, queue,
r1 -> new Thread(r1, "MatsFuturizer completer #" + _threadNumber.getAndIncrement()));
/*
* Part 2: We make a special RejectionExecutionHandler which upon rejection due to "full queue" (i.e. TPE has
* reached max pool size) puts the task on queue anyway, using our sneaky method (LTQ is not bounded).
*/
threadPool.setRejectedExecutionHandler((r, executor) -> {
queue.sneak(r);
});
return threadPool;
}
protected Promise _createPromise(String traceId, String from, String to, Class replyClass,
int timeout, TimeUnit unit) {
long timeoutMillis = unit.toMillis(timeout);
if (timeoutMillis <= 0) {
throw new IllegalArgumentException("Timeout in milliseconds cannot be zero or negative [" + timeoutMillis
+ "].");
}
String correlationId = RandomString.randomCorrelationId();
long timestamp = System.currentTimeMillis();
CompletableFuture> future = new CompletableFuture<>();
if (log.isDebugEnabled()) log.debug(LOG_PREFIX + "Creating Promise for TraceId [" + traceId + "], from [" + from
+ "], to [" + to + "], timeout in [" + timeoutMillis + "] millis.");
return new Promise<>(traceId, correlationId, from, to, timestamp, System.nanoTime(), timestamp + timeoutMillis,
replyClass, future);
}
protected void _enqueuePromise(Promise promise) {
_internalStateLock.lock();
try {
if (_correlationIdToPromiseMap.size() >= _maxOutstandingPromises) {
throw new IllegalStateException("There are too many Promises outstanding, so cannot add more"
+ " - limit is [" + _maxOutstandingPromises + "].");
}
// This is the lookup that the reply-handler uses to get to the promise from the correlationId.
_correlationIdToPromiseMap.put(promise._correlationId, promise);
// This is the priority queue that the timeouter-thread uses to get the next Promise to timeout.
_timeoutSortedPromises.add(promise);
// ?: Have the earliest Promise to timeout changed by adding this Promise?
if (_nextInLineToTimeout != _timeoutSortedPromises.peek()) {
// -> Yes, this was evidently earlier than the one we had "next in line", so notify the timeouter-thread
// that a new promise was entered, to re-evaluate "next to timeout".
_timeouterPing_InternalStateLock.signal();
}
}
finally {
_internalStateLock.unlock();
}
}
protected volatile boolean _replyHandlerEndpointStarted;
protected void _assertFuturizerRunning() {
// ?: Have we already checked that the reply endpoint is running?
if (!_replyHandlerEndpointStarted) {
// -> No, so wait for it to start now
boolean started = _replyHandlerEndpoint.waitForReceiving(60_000);
// ?: Did it start?
if (!started) {
// -> No, so that's bad.
throw new IllegalStateException("The Reply Handler SubscriptionTerminator Endpoint would not start.");
}
// Shortcut this question forever after.
_replyHandlerEndpointStarted = true;
}
// ?: Have we already shut down?
if (!_runFlag) {
// -> Yes, shut down, so that's bad.
throw new IllegalStateException("This MatsFuturizer [" + _terminatorEndpointId + "] is shut down.");
}
}
protected void _sendRequestToFulfillPromise(String from, String endpointId, String traceId, Object request,
InitiateLambda extraMessageInit, Promise promise) {
_matsInitiator.initiateUnchecked(msg -> {
// Stash in the standard stuff
msg.traceId(traceId)
.from(from)
.to(endpointId)
.replyToSubscription(_terminatorEndpointId, promise._correlationId);
// Stash up with any extra initialization stuff
extraMessageInit.initiate(msg);
// Do the request.
msg.request(request);
});
}
protected final ReentrantLock _internalStateLock = new ReentrantLock();
protected final Condition _timeouterPing_InternalStateLock = _internalStateLock.newCondition();
// Synchronized on _internalStateLock
protected final HashMap> _correlationIdToPromiseMap = new HashMap<>();
// Synchronized on _internalStateLock
protected final PriorityQueue> _timeoutSortedPromises = new PriorityQueue<>();
// Synchronized on _internalStateLock
protected Promise _nextInLineToTimeout;
protected void _handleRepliesForPromises(ProcessContext context, String correlationId,
MatsObject matsObject) {
// Immediately pick this out of the map & queue
Promise promise;
_internalStateLock.lock();
try {
// Find the Promise from the CorrelationId
promise = _correlationIdToPromiseMap.remove(correlationId);
// Did we find it?
if (promise != null) {
// -> Yes, found - remove it from the PriorityQueue too.
_timeoutSortedPromises.remove(promise);
}
// NOTE: We don't bother pinging the Timeouter, as he'll find out himself soon enough if this was first.
}
finally {
_internalStateLock.unlock();
}
// ?: Did we still have the Promise?
if (promise == null) {
// -> Promise gone, log on INFO and exit (it was logged on WARN when it was actually timed out).
MDC.put("traceId", context.getTraceId());
log.info(LOG_PREFIX + "Promise gone! Got reply from [" + context
.getFromStageId() + "] for Future with traceId:[" + context.getTraceId()
+ "], but the Promise had timed out.");
MDC.remove("traceId");
return;
}
// ----- We have Promise, and shall now fulfill it. Send off to pool thread.
_futureCompleterThreadPool.execute(() -> {
try {
MDC.put(MDC_TRACE_ID, promise._traceId);
MDC.put(MDC_MATS_INIT_ID, promise._from);
// NOTICE! We don't log here, as the SubscriptionTerminator already has logged the ordinary mats lines.
if (log.isDebugEnabled()) log.debug(LOG_PREFIX + "Completing promise from [" + promise._from + "]: ["
+ promise + "]");
Object replyObject;
try {
replyObject = _deserializeReply(matsObject, promise._replyClass);
}
catch (Throwable t) { // Notice: Unless overridden, this should always be IllegalArgumentException.
log.error("Got problems completing Future due to failing to deserialize the incoming object to"
+ " expected class [" + promise._replyClass.getName() + "], thus doing"
+ " future.completeExceptionally(..) with the [" + t.getClass().getSimpleName() + "]."
+ " Initiated from [" + promise._from + "], with reply from [" + context.getFromStageId()
+ "], traceId [" + context.getTraceId() + "]", t);
promise._future.completeExceptionally(t);
return;
}
_completeFuture(context, replyObject, promise);
}
// NOTICE! This catch will probably never be triggered, as if .thenAccept() and similar throws,
// the CompletableFuture evidently handles it and completes the future exceptionally.
catch (Throwable t) {
log.error(LOG_PREFIX + "Got problems completing Future initiated from [" + promise._from
+ "], with reply from [" + context.getFromStageId()
+ "], traceId:[" + context.getTraceId() + "]", t);
}
finally {
// This is a MatsFuturizer thread pool thread, so we own it. Clear MDC.
MDC.clear();
}
});
}
protected Object _deserializeReply(MatsObject matsObject, Class toClass) {
return matsObject.toClass(toClass);
}
private static final Logger log_reply = LoggerFactory.getLogger(MatsFuturizer.class.getName() + ".Reply");
@SuppressWarnings({ "unchecked", "rawtypes" }) // We know that the futureReply is of the same type as the Promise.
protected void _completeFuture(ProcessContext context, Object replyObject, Promise promise) {
Reply futureReply = new Reply<>(context, replyObject, promise._initiationTimestamp,
promise._initiationNanos);
// If special Reply-logger is INFO-enabled, log a line when the getter is invoked.
// ?: Is the logger enabled?
if (log_reply.isInfoEnabled()) {
// -> Yes, logger enabled, so time the future completion, fill the MDC and log a line.
long nanosAtStart_completing = System.nanoTime();
// ::: === Actual Future.complete(..)!
promise._future.complete((Reply) futureReply);
long nanosNow = System.nanoTime();
long nanosTaken_completing = nanosNow - nanosAtStart_completing;
long nanosTaken_total = nanosNow - promise._initiationNanos;
// Microseconds should be plenty resolution.
double roundTripMillis = Math.round(futureReply.roundTripNanos / 1000d) / 1000d;
double completingMillis = Math.round(nanosTaken_completing / 1000d) / 1000d;
double totalMillis = Math.round(nanosTaken_total / 1000d) / 1000d;
MDC.put(MDC_MATS_FUTURE_COMPLETED, Double.toString(totalMillis));
MDC.put(MDC_MATS_FUTURE_TIME_RTT, Double.toString(roundTripMillis));
MDC.put(MDC_MATS_FUTURE_TIME_COMPLETING, Double.toString(completingMillis));
// NOTICE: No need to clean MDC, as it is _cleared_ by caller after this method returns.
log_reply.info(MatsFuturizer.LOG_PREFIX + "Completed Future from initiatorId"
+ " [" + promise._from + "] with answer from [" + context.getFromStageId()
+ (replyObject != null
? "], with instance of [" + replyObject.getClass().getSimpleName() + "]"
: "], which was null")
+ " - Total:[" + totalMillis + " ms], Mats RTT:[" + roundTripMillis + " ms].");
}
else {
// -> No, logger not enabled, so don't bother timing the future completion either.
// ::: === Actual Future.complete(..)!
promise._future.complete((Reply) futureReply);
}
}
protected volatile boolean _runFlag = true;
protected void _startTimeouterThread() {
Runnable timeouter = () -> {
log.info(LOG_PREFIX + "MatsFuturizer Timeouter-thread: Started!");
while (_runFlag) {
List> promisesToTimeout = new ArrayList<>();
_internalStateLock.lock();
try {
while (_runFlag) {
try {
long sleepMillis;
long now = System.currentTimeMillis();
Promise peekPromise = _timeoutSortedPromises.peek();
if (peekPromise != null) {
// ?: Is this Promise overdue? I.e. current time has passed timeout timestamp of
// promise.
if (now >= peekPromise._timeoutTimestamp) {
// -> Yes, timed out. remove from both collections
if (log.isDebugEnabled()) log.debug(LOG_PREFIX + "Promise at head of timeout queue"
+ " HAS timed out [" + (now - peekPromise._timeoutTimestamp)
+ "] millis ago - traceId [" + peekPromise._traceId + "].");
// It is the first, since it is the object we peeked at.
_timeoutSortedPromises.remove();
// Remove explicitly by CorrelationId.
_correlationIdToPromiseMap.remove(peekPromise._correlationId);
// Put it in the list to timeout
promisesToTimeout.add(peekPromise);
// Check next in line
continue;
}
// E-> This is the Promise that is next in line to timeout.
_nextInLineToTimeout = peekPromise;
// This Promise has >0 milliseconds left before timeout, so calculate how long to sleep.
sleepMillis = peekPromise._timeoutTimestamp - now;
if (log.isDebugEnabled()) log.debug(LOG_PREFIX + "Promise at head of timeout queue has"
+ " NOT timed out, will time out in [" + sleepMillis + "] millis - traceId ["
+ peekPromise._traceId + "].");
}
else {
// We have no Promise next in line to timeout.
// Note: NOT logging to NOT be annoying in a dev situation.
_nextInLineToTimeout = null;
// Sleep forever until notified, where "forever" means 30 seconds - before checking
// again to be sure..!
sleepMillis = 30_000;
}
// ?: Did we find any Promises to timeout?
if (!promisesToTimeout.isEmpty()) {
// -> Yes, Promises to timeout - exit out of synch and inner run-loop to do that.
break;
}
// ----- We've found a new sleep time, go sleep.
// :: Now go to sleep, waiting for signal from "new element added" or close()
long nanosStart_sleep = 0;
// ?: Is debug enabled AND we actually have a Promise we're sleeping for.
if (log.isDebugEnabled() && (_nextInLineToTimeout != null)) {
nanosStart_sleep = System.nanoTime();
log.debug(LOG_PREFIX + "Will now go to sleep for [" + sleepMillis + "] millis.");
}
// Do the sleep (.. which is a Condition.await(..) on the _internalStateLock)
_timeouterPing_InternalStateLock.await(sleepMillis, TimeUnit.MILLISECONDS);
if (log.isDebugEnabled() && (_nextInLineToTimeout != null)) {
double millisSlept = (System.nanoTime() - nanosStart_sleep) / 1_000_000d;
log.debug(LOG_PREFIX + ".. slept [" + millisSlept + "] millis (should have slept ["
+ sleepMillis + "] millis, difference [" + (millisSlept - sleepMillis)
+ "] millis too much).");
}
}
// :: Protection against bad code - catch-all Throwables in hope that it will auto-correct.
catch (Throwable t) {
log.error(LOG_PREFIX + "Got an unexpected Throwable in the promise-timeouter-thread."
+ " Loop and check whether to exit.", t);
// If exiting, do it now.
if (!_runFlag) {
break;
}
// :: Protection against bad code - sleep a tad to not tight-loop.
try {
Thread.sleep(10_000);
}
catch (InterruptedException e) {
e.printStackTrace();
}
}
}
}
finally {
_internalStateLock.unlock();
}
// ----- This is outside the synch block
// :: Timing out Promises that was found to be overdue.
int promisesToTimeoutCount = promisesToTimeout.size();
if (log.isDebugEnabled()) log.debug(LOG_PREFIX + "Will now timeout [" + promisesToTimeoutCount
+ "] Promise(s).");
for (Promise promise : promisesToTimeout) {
_futureCompleterThreadPool.execute(() -> {
try {
double millisSinceInitiation = Math.round((System.nanoTime() - promise._initiationNanos)
/ 1000d) / 1000d;
MDC.put(MDC_TRACE_ID, promise._traceId);
MDC.put(MDC_MATS_INIT_ID, promise._from);
MDC.put(MDC_MATS_FUTURE_TIMEOUT, Double.toString(millisSinceInitiation));
String msg = "The Promise/Future timed out! It was initiated from:[" + promise._from
+ "] with traceId:[" + promise._traceId + "], to:[" + promise._to + "]"
+ " Initiation was [" + millisSinceInitiation + " ms] ago, and its specified"
+ " timeout was:[" + (promise._timeoutTimestamp - promise._initiationTimestamp)
+ "].";
log.warn(LOG_PREFIX + msg);
// Timeout
_timeoutCompleteExceptionally(promise, msg);
}
// NOTICE! This catch will probably never be triggered, as if .thenAccept() and similar throws,
// the CompletableFuture evidently handles it and completes the future exceptionally.
catch (Throwable t) {
log.error(LOG_PREFIX + "Got problems timing out Promise/Future initiated from:["
+ promise._from + "] with traceId:[" + promise._traceId + "], ignoring.", t);
}
finally {
// This is a MatsFuturizer thread pool thread, so we own it. Clear MDC.
MDC.clear();
}
});
// This is a MatsFuturizer timeouter-thread, so we own it. Clear MDC.
MDC.clear();
/*
* Wild hack to get unit tests to pass on annoying MacOS: Both Object.wait(..), and
* ReentrantLock.newCondition().await(..) gives wildly bad oversleeping on MacOS, in excess of 200
* ms (in contrast, my Linux box is consistenly <0.2 ms off). Thus, when submitting multiple futures
* with timeout spaced 100 ms apart (as in Test_MatsFuturizer_Timeouts), we sometimes end up timing
* out multiple futures in one go. However, these are still timed out in the correct order. They
* would thus have come back to the test with the correct order, was it not for the moving over to
* the _futureCompleterThreadPool - where these "double" timeoutings might change order. Therefore,
* if there are <1, 10> futures to timeout, we'll sleep a small while between each.
*/
if ((promisesToTimeoutCount > 1) && (promisesToTimeoutCount < 10)) {
try {
Thread.sleep(5);
}
catch (InterruptedException e) {
/* Ignore, as we'll check the runFlag-condition in the loop. */
}
}
}
promisesToTimeout.clear();
// .. will now loop into the synch block again.
}
log.info("MatsFuturizer Timeouter-thread: We got asked to exit, and that we do!");
};
new Thread(timeouter, "MatsFuturizer Timeouter").start();
}
protected void _timeoutCompleteExceptionally(Promise promise, String msg) {
promise._future.completeExceptionally(new MatsFuturizerTimeoutException(
msg, promise._initiationTimestamp, promise._traceId));
}
/**
* Closes the MatsFuturizer. Notice: Spring will also notice this method if the MatsFuturizer is registered as a
* @Bean, and will register it as a destroy method.
*/
public void close() {
if (!_runFlag) {
log.info("MatsFuturizer.close() invoked, but runFlag is already false, thus it has already been closed.");
return;
}
log.info("MatsFuturizer.close() invoked: Shutting down & removing reply-handler-endpoint,"
+ " shutting down future-completer-threadpool, timeouter-thread,"
+ " and cancelling any outstanding futures.");
_runFlag = false;
_replyHandlerEndpoint.remove(5000);
_futureCompleterThreadPool.shutdown();
// :: Find all remaining Promises, and notify Timeouter-thread that we're dead.
List> promisesToCancel = new ArrayList<>();
_internalStateLock.lock();
try {
promisesToCancel.addAll(_timeoutSortedPromises);
// Clear the collections, just to have a clear conscience.
_timeoutSortedPromises.clear();
_correlationIdToPromiseMap.clear();
// Notify the Timeouter-thread that shit is going down.
_timeouterPing_InternalStateLock.signalAll();
}
finally {
_internalStateLock.unlock();
}
// :: Cancel all outstanding Promises.
for (Promise promise : promisesToCancel) {
try {
MDC.put("traceId", promise._traceId);
promise._future.cancel(true);
}
// NOTICE! This catch will probably never be triggered, as if .thenAccept() and similar throws,
// the CompletableFuture evidently handles it and completes the future exceptionally.
catch (Throwable t) {
log.error(LOG_PREFIX + "Got problems cancelling (due to shutdown) Promise/Future initiated from:["
+ promise._from + "] with traceId:[" + promise._traceId + "]", t);
}
finally {
MDC.remove("traceId");
}
}
}
}
© 2015 - 2024 Weber Informatics LLC | Privacy Policy