[Bug]: telemetry truncation time grows with the square of the operation count for size-limited exporters other than LambdaLogExporter

Ouverte
  • Quelque chose ne va pas ?
#877 0 commentaires 0 réactions 0 personnes assignées Voir sur GitHub

Les mainteneurs répondent en général sous 1 jour

Personne n'a encore pris cette issue.

Évaluation

Difficulté
4/5
Temps estimé
3-5 jours
Accessibilité débutants
54/100
Type d'issue
Bug
Clarté
Clairement spécifiée
Activité
Active
Stack technique
aws, java
Domaine
backend, performance

Piste de recherche

Start with telemetry-exporter/src/main/java/software/amazon/lambda/durable/telemetry/Truncation.java, especially truncateRecord and Meter.fits, then trace how DurableExecutionTelemetry.exportRecord chooses the truncation path. Compare the built-in exporter renderings and the bounded meter behavior described in the issue. Done means oversized records truncate without the reported repeated full-record measurement cost while preserving the record format and correct truncation results.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Description

bug pkg:telemetry
Expected Behavior

For every built-in size-limited exporter, truncating an oversized record takes a small share of the invocation, whatever the number of operations.

Actual Behavior

DurableExecutionTelemetry.exportRecord truncates with Truncation.truncateMeasured for LambdaLogExporter, and with Truncation.truncateRecord for every other exporter. truncateRecord passes bounded = false. So when a rendered record is over its limit, Meter.fits measures it before it considers each operation, once in the result step and again in the operation step. The work grows with the square of the operation count. A record that already fits returns after one measurement.

With the limit and rendering of CloudWatchLogsExporter (256,000 bytes, operations by name), truncating a record of 4,000 simple successful steps took 15.4 s on JDK 21.0.13 at tag v3.0.1. With the same limit and the array rendering, it took 3.5 s for 2,000 operations and 14.8 s for 4,000. The LambdaLogExporter path truncated a 30,000-operation record in 0.6 s.

Every other built-in exporter with a default limit takes the slow path when a record is over its limit: CloudWatchLogsExporter, SQSExporter and EventBridgeExporter (256,000 bytes), DynamoDBExporter (400,000), FirehoseExporter, AuroraExporter, RedshiftExporter and OTelExporter (1,000,000), S3Exporter (5,000,000) and OpenSearchExporter (10,000,000). FileExporter and HttpExporter take it too when a limit is configured.

The time is spent before the invocation returns:

  • TelemetryPlugin.onInvocationEnd waits for the export through ExportScheduler.drain, which has no deadline.
  • DurableExecutor calls onInvocationEnd before it returns the invocation output.
  • The final record also includes operations from earlier invocations. With the default settings it lists every named top-level operation of the execution. So a long execution truncates its accumulated operations when it ends.
  • With EmitMode.ON_CHANGE, the plugin also exports records during the invocation.
Steps to Reproduce

Check out tag v3.0.1. Add this test to telemetry-exporter/src/test/java/software/amazon/lambda/durable/telemetry/, then run mvn -pl telemetry-exporter -am -Djacoco.skip=true -Dtest=TruncationTimeTest -Dsurefire.failIfNoSpecifiedTests=false test:

package software.amazon.lambda.durable.telemetry;

import java.time.Instant;
import java.time.ZoneOffset;
import java.time.format.DateTimeFormatter;
import org.junit.jupiter.api.Test;

class TruncationTimeTest {
    private static final DateTimeFormatter TS =
            DateTimeFormatter.ofPattern("yyyy-MM-dd'T'HH:mm:ss.SSS'Z'").withZone(ZoneOffset.UTC);

    @Test
    void truncateRecordTime() {
        DurableExecutionTelemetryRecord record = new DurableExecutionTelemetryRecord();
        record.executionArn =
                "arn:aws:lambda:us-east-1:123456789012:function:demo:$LATEST/durable-execution/run-1/inv-1";
        record.status = "SUCCEEDED";
        Instant start = Instant.parse("2026-10-28T00:00:00Z");
        for (int i = 0; i < 4000; i++) {
            Instant s = start.plusMillis(i * 10L);
            record.operations.add(new OperationRecord()
                    .id("op-" + i)
                    .name("step-" + i)
                    .type("STEP")
                    .subType("Step")
                    .status("SUCCEEDED")
                    .startTime(TS.format(s))
                    .endTime(TS.format(s.plusMillis(3)))
                    .durationMs(3L)
                    .attempt(1));
        }

        // CloudWatchLogsExporter's limit and rendering: 256,000 bytes, operations by name.
        long t0 = System.nanoTime();
        DurableExecutionTelemetryRecord out =
                Truncation.truncateRecord(record, 256_000, DurableExecutionTelemetryRecord::toByNameWireMap);
        System.out.printf(
                "truncateRecord: %.1f s, kept %d operations, dropped %d%n",
                (System.nanoTime() - t0) / 1e9, out.operations().size(), out.droppedOperations);
    }
}

It prints a time of about 15 s and kept 1568 operations, dropped 2432.

SDK Version

aws-durable-execution-sdk-java-plugin-telemetry-exporter 3.0.1. main has the same truncation code.

Java Version

21

Is this a regression?

No

Additional Context

Two changes would remove most of the cost. Neither changes the public API or the record format.

  1. Measure once after the error causes are removed. Then, in the result step, measure again only after an operation loses its result. With the default content settings no operation has a result. So today that step measures the full record once per operation and drops nothing.
  2. Use the bounded meter for the built-in renderings whose per-drop shrink has a known upper bound. truncateMeasured documents that bound for the record's JSON and for that JSON embedded as one string value. The array and by-name wire maps fit it, and so does the OTel body, which embeds the rendered record as a string. OperationsFormat.BOTH holds the operations array and the derived by-name map, so its bound must cover the shrink of both. A custom render keeps exact checks. For a rendering that the bound covers, the bounded meter returns the same record as checking after every drop.

The Go SDK had the same problem (Go SDK #153: Truncate re-renders the whole record per dropped operation) and fixed it with a binary search over the drop count (Go SDK PR #154: binary-search the Truncate drop count).

Code at the commit I read:

Langage dominant
Java
Étoiles
28
Forks
18
Merge moyen
18 h 16 min
PR mergées (30 j)
62

Préparer son environnement

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de aws/aws-durable-execution-sdk-java

Toutes les issues de aws/aws-durable-execution-sdk-java

Issues similaires

Plus d'issues Java

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.