问题
I'm evaluating Esper as a system for loss-less processing of billing data. It is expected that system can handle ~20000 events per second and run ~400 statements with continuos aggregation (without storing events in memory). In order to get expected performance I've started to send events in multiple threads and found that esper often looses data.
Simple example that shows data loss
import java.util.ArrayList;
import java.util.Arrays;
import java.util.Collections;
import java.util.List;
import java.util.concurrent.CompletableFuture;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.TimeUnit;
import java.util.stream.Collectors;
import com.espertech.esper.client.Configuration;
import com.espertech.esper.client.EPAdministrator;
import com.espertech.esper.client.EPRuntime;
import com.espertech.esper.client.EPServiceProvider;
import com.espertech.esper.client.EPServiceProviderManager;
import com.espertech.esper.client.EPStatement;
public class Example {
public static void main(String[] args) throws Exception {
new Example().run();
}
public void run() throws Exception {
Configuration config = new Configuration();
// use default configuration
EPServiceProvider epService = EPServiceProviderManager.getDefaultProvider(config);
EPAdministrator epAdministrator = epService.getEPAdministrator();
// simple schema
epAdministrator.getConfiguration().addEventType(LogLine.class);
// event for terminating context partition
createEPL(epAdministrator, "create schema TerminateEvent() ");
// Start context partition on LogLine event and terminate on TerminateEvent.
createEPL(epAdministrator, "create context InitCtx start LogLine end TerminateEvent");
// select to collect count of events per account_name.
EPStatement statement = createEPL(epAdministrator, "context InitCtx select context.id as partition_id, count(*), sum(bytes) from LogLine output last when terminated");
// register listener to output all newEvents properties values
statement.addListener((newEvents, oldEvents) -> {
String resultEvents = Arrays.stream(newEvents).map((event) -> {
return Arrays.stream(event.getEventType().getPropertyNames()).map((prop) -> {
return prop + "=" + event.get(prop);
}).collect(Collectors.joining(", "));
}).collect(Collectors.joining("]; ["));
System.out.println("=== results: [" + resultEvents + "]");
});
//lets use 4 threads for sending data
ExecutorService myexecutor = Executors.newFixedThreadPool(4);
List<CompletableFuture<Void>> listOfTasks = new ArrayList<>();
//get data to be processed
List<LogLine> list = getData();
for (int i = 1; i <= list.size(); i++) {
//concurrently send each logline
final LogLine logLine = list.get(i - 1);
CompletableFuture<Void> task = CompletableFuture.runAsync(() -> {
epService.getEPRuntime().sendEvent(logLine);
System.out.println("== sending data " + logLine);
}, myexecutor);
listOfTasks.add(task);
if (i % 4 == 0) {
// terminate context partition after every 4 events.
sendTerminateEvent(listOfTasks, epService.getEPRuntime());
}
}
// terminate context partition at the end of the execution.
sendTerminateEvent(listOfTasks, epService.getEPRuntime());
// shutdow all services.
myexecutor.shutdown();
epService.destroy();
}
private void sendTerminateEvent(List<CompletableFuture<Void>> listOfTasks, EPRuntime epRuntime) throws Exception {
// wait for all submitted tasks to finish
CompletableFuture[] array = listOfTasks.toArray(new CompletableFuture[listOfTasks.size()]);
CompletableFuture.allOf(array).get(1, TimeUnit.MINUTES);
listOfTasks.clear();
System.out.println("== sending terminate event.");
// send partition termination event
epRuntime.sendEvent(Collections.emptyMap(), "TerminateEvent");
}
private List<LogLine> getData() {
List<LogLine> dataEventsList = new ArrayList<>();
dataEventsList.add(new LogLine(0, 1));
dataEventsList.add(new LogLine(0, 2));
dataEventsList.add(new LogLine(0, 3));
dataEventsList.add(new LogLine(0, 4));
dataEventsList.add(new LogLine(0, 5));
dataEventsList.add(new LogLine(1, 1));
dataEventsList.add(new LogLine(1, 2));
dataEventsList.add(new LogLine(1, 3));
dataEventsList.add(new LogLine(1, 4));
dataEventsList.add(new LogLine(1, 5));
return dataEventsList;
}
private EPStatement createEPL(EPAdministrator admin, String statement) {
System.out.println("creating EPL: " + statement);
return admin.createEPL(statement);
}
public static class LogLine {
int account_id;
int bytes;
public LogLine(int account_id, int bytes) {
this.account_id = account_id;
this.bytes = bytes;
}
public int getAccount_id() {
return account_id;
}
public int getBytes() {
return bytes;
}
@Override
public String toString() {
return "[account_id=" + account_id + ", bytes=" + bytes + "]";
}
}
}
Execution output:
creating EPL: create schema TerminateEvent()
creating EPL: create context InitCtx start LogLine end TerminateEvent
creating EPL: context InitCtx select context.id as partition_id, count(*), sum(bytes) from LogLine output last when terminated
== data [account_id=0, bytes=3] was send
== data [account_id=0, bytes=1] was send
== data [account_id=0, bytes=4] was send
== data [account_id=0, bytes=2] was send
== sending terminate event.
=== results: [partition_id=0, count(*)=4, sum(bytes)=10]
== data [account_id=1, bytes=2] was send
== data [account_id=1, bytes=3] was send
== data [account_id=0, bytes=5] was send
== data [account_id=1, bytes=1] was send
== sending terminate event.
=== results: [partition_id=1, count(*)=2, sum(bytes)=6]
== data [account_id=1, bytes=5] was send
== data [account_id=1, bytes=4] was send
== sending terminate event.
=== results: [partition_id=2, count(*)=1, sum(bytes)=4]
There are correct results for the first partition, the next 2 partitions output invalid results:
// OK
actual [partition_id=0, count(*)=4, sum(bytes)=10]
expected [partition_id=0, count(*)=4, sum(bytes)=10]
// LOSS
actual [partition_id=1, count(*)=2, sum(bytes)=6]
expected [partition_id=1, count(*)=4, sum(bytes)=11]
// LOSS
actual [partition_id=2, count(*)=1, sum(bytes)=4]
expected [partition_id=2, count(*)=2, sum(bytes)=9]
What's wrong with this example code?
Enabling priority execution order didn't help
creating EPL: create schema TerminateEvent()
creating EPL: @Priority(1) create context InitCtx start LogLine end TerminateEvent
creating EPL: @Priority(0) context InitCtx select context.id as partition_id, count(*), sum(bytes) from LogLine output last when terminated
== data [account_id=0, bytes=3] was send
== data [account_id=0, bytes=4] was send
== data [account_id=0, bytes=1] was send
== data [account_id=0, bytes=2] was send
== sending terminate event.
=== results: [partition_id=0, count(*)=4, sum(bytes)=10]
== data [account_id=1, bytes=2] was send
== data [account_id=1, bytes=3] was send
== data [account_id=0, bytes=5] was send
== data [account_id=1, bytes=1] was send
== sending terminate event.
=== results: [partition_id=1, count(*)=2, sum(bytes)=6]
== data [account_id=1, bytes=5] was send
== data [account_id=1, bytes=4] was send
== sending terminate event.
=== results: [partition_id=2, count(*)=1, sum(bytes)=4]
回答1:
This question is a more elaborate DUPLICATE of Esper data loss when inbound threading is enabled
In the case that Esper EPL requires ordered execution you must develop your code so it processes events in an ordered fashion. Esper cannot magically enforce some ordering. The JVM can pause any thread at any time. You must design your code properly.
For example, lets assume you have 2 threads. Lets assume that A can be processed in parallel and B must be processed exactly in the order as provided in the example below.
Lets say you have events come in. You want B to process after A1 and A2 but before A3 and A4:
A1 A2 B1 A3 A4
If you just add all A and B events to a queue and thread pool with say 5 threads that means that B can be processed first, in the middle, or last. Every run you can get a different result as the JVM does not enforce an order. Esper cannot enforce an order as your application drives Esper and not the other way around.
What you can do, for example, is add the first set of A events to a queue (A1, A2). When B comes in, wait for the queue to empty. Next add B to the queue. Wait for B to complete. Then add the next set of A events (A3, A4) into the queue. Thus you achieve ordered processing in respect to A and B, with all A events being processed in parallel.
CORRECTION:
I see now that you only have one event type and no A+B. In this case make sure you are running the most recent version. Also make sure that "create context" does not get a lower priority otherwise context partitions get created last. I have run your code around 10 times and did not see invalid output with 7.1.0. I'm on JDK 1.8.0_121 (Oracle).
来源：https://stackoverflow.com/questions/49738825/esper-loss-less-events-processing