Throughput Tuning

Small, release-backed Elsa 3.8.0 configuration examples for testing commit, worker, and transactional-outbox tradeoffs under load.

These are starting configurations, not production targets. Benchmark each change against your own workflow shape, persistence provider, and downstream services. Keep recovery tests in the benchmark: throughput is only useful when the workflow remains correct after a restart or fault.

Register selectable commit strategies

This configuration registers the standard strategy names and a named periodic strategy. A workflow definition can select any registered name through WorkflowOptions.CommitStrategyName or the Elsa Studio Properties → Settings → Commit Strategy selector.

using Elsa.Extensions;
using Elsa.Workflows.CommitStates;
using Elsa.Workflows.CommitStates.Strategies;

builder.Services.AddElsa(elsa =>
{
    elsa.UseWorkflows(workflows =>
    {
        workflows.UseCommitStrategies(strategies =>
        {
            strategies.AddStandardStrategies();
            strategies.Add(
                "Periodic30Seconds",
                "Every 30 seconds",
                "Commit workflow state at least every 30 seconds during execution.",
                new PeriodicWorkflowStrategy(TimeSpan.FromSeconds(30)));
        });
    });
});

Use WorkflowExecuted as a test candidate for short workflows only after verifying the state that must survive a failure before completion. Use ActivityExecuted when persistence after each completed activity is more important than the additional writes. Test a named periodic strategy for long-running workflows that need a bounded recovery interval. In this example, the stored strategy name is Periodic30Seconds; Studio shows Every 30 seconds.

Set a host fallback

Definitions with no commit-strategy selection use the host fallback. This is not added to the Studio/API registry, so register a named strategy separately when authors should be able to choose it.

using Elsa.Extensions;
using Elsa.Workflows.CommitStates.Strategies;

builder.Services.AddElsa(elsa =>
{
    elsa.UseWorkflows(workflows =>
    {
        workflows.WithDefaultWorkflowCommitStrategy(
            new PeriodicWorkflowStrategy(TimeSpan.FromSeconds(30)));
    });
});

Increase only the measured worker queue

The three mediator queues default to four workers each. If command processing is the proven bottleneck, begin with a modest command-only change and compare the result with the baseline.

using Elsa.Mediator.Options;

builder.Services.Configure<MediatorOptions>(options =>
{
    options.CommandWorkerCount = 8;
    options.JobWorkerCount = 4;
    options.NotificationWorkerCount = 4;
});

The command path is relevant to background workflow dispatch. Do not increase the other counts merely to match it: notifications and jobs can put additional load on the same persistence and external dependencies.

Use the transactional outbox for in-workflow dispatch

Enable the outbox when dispatching another workflow must be coordinated with the current workflow's state commit. It is not a free throughput feature: it adds persisted outbox work, so measure both successful dispatch latency and recovery behavior.

using Elsa.Workflows.Runtime.Options;

builder.Services.Configure<WorkflowDispatcherOptions>(options =>
{
    options.UseTransactionalOutbox = true;
    options.ProcessOutboxAfterCommit = true;
    options.OutboxProcessorBatchSize = 100;
});

If the immediate delivery work is itself a measurable commit-path bottleneck, test ProcessOutboxAfterCommit = false. Delivery will then rely on the recurring processor, so validate the resulting dispatch delay and restart behavior before adopting it.

Compare results consistently

For every run, capture:

• completed workflows per interval and end-to-end latency;

• elsa.workflow.started, elsa.workflow.completed, and elsa.workflow.faulted rates;

• elsa.activity.duration percentiles and the slowest activity spans;

• database resource use and commit latency; and

• the result of a forced-restart recovery test.

Use the same traffic mix for each run. Change one variable, retain the result only if it improves the system rather than shifting pressure to the database or a downstream dependency.

See Performance tuning for the decision guide and Distributed tracing for Elsa's OpenTelemetry instrumentation.