Monitoring And Metrics

This guide covers CloudWatch metrics analysis and custom metrics emission for Amazon Managed Service for Apache Flink applications. Use it when setting up monitoring for a new application, interpreting metric behavior on an existing one, or adding application-specific business metrics via the Flink MetricGroup API.

Key CloudWatch Metrics for Managed Service for Apache Flink

Core Metrics Reference

Managed Service for Apache Flink publishes metrics to CloudWatch under the AWS/KinesisAnalytics namespace. The following table lists the key metrics, what they measure, and expected healthy ranges for a well-tuned application.

Throughput Metrics

These metrics are available at Application, Operator, Task, and Parallelism levels. They have a special reporting behavior: Managed Service for Apache Flink takes 4 metric snapshots per minute. When using the SUM statistic over a period, you must divide by 4 to get the correct value (metric math: m1/4).

Important: Select the metric at the correct Level. If you’re tracking the metric for an Operator, you need to select the corresponding operator-level metric, not the application-level one.

Operator Performance Metrics

These metrics are available at Task, Operator, and Parallelism levels. They are available for Flink version 1.13 and later.

Watermark Metrics

Managed Memory Metrics (Flink 1.13+)

These metrics relate to memory managed by Flink outside the Java heap, used for the RocksDB state backend.

Kinesis Source Metrics

Kafka Source Metrics

How cpuUtilization and heapMemoryUtilization Are Reported

A common source of confusion: these metrics are not reported as a single summed value across all TaskManagers. Instead, Managed Service for Apache Flink publishes one sample per TaskManager per reporting interval.

With 5 TaskManagers, CloudWatch receives 5 data points per interval. CloudWatch statistics then work naturally:

• Average = mean utilization across all TaskManagers
• Maximum = the hottest (most utilized) TaskManager
• Minimum = the least utilized TaskManager

You do not need to divide by KPU count. Use Maximum to find the bottleneck TaskManager, or Average for overall health. The AWS docs recommend using Maximum for alarm thresholds.

Note: containerCPUUtilization and containerMemoryUtilization publish 2× the number of samples per TaskManager per minute (2 × number of TaskManagers).

Metric Dimensions

Managed Service for Apache Flink metrics are available at four dimension levels:

Dimension selection guidance:

• Start with Application-level metrics for dashboards and alarms — they give the overall health picture with minimal cardinality.
• Drop to Task-level when Application-level metrics show a problem (e.g., high backpressure) and you need to identify which task chain is responsible.
• Use Operator-level for targeted debugging of specific operators.
• Use Parallelism-level only for diagnosing data skew or per-shard/partition analysis, not for persistent dashboards.

Warning: Task-Level Metrics and High Cardinality

Task-level, operator-level, and parallelism-level metrics produce one CloudWatch metric per subtask per metric name. The total metric count is:

With high parallelism and ParallelismPerKPU > 1, this can produce thousands of individual CloudWatch metrics, leading to:

• High CloudWatch costs: CloudWatch charges per metric per month. An application with Parallelism = 64 (32 KPUs × ParallelismPerKPU = 2) × 50 operator-level metrics = 3,200 custom metrics.
• Dashboard performance degradation: CloudWatch dashboards with thousands of metrics load slowly and become difficult to navigate.
• Alarm limits: CloudWatch has per-account alarm limits. Task-level alarms at high parallelism can consume a significant portion of the quota.

Recommendation: Use Application-level dimensions for all persistent alarms and dashboards. Reserve Task-level, Operator-level, and Parallelism-level dimensions for temporary diagnostic queries using CloudWatch Metrics Insights or the Flink Web UI, which provides subtask-level metrics without CloudWatch cardinality costs.

Recommended Alarms and Thresholds

Critical Metric Alarms

Set up CloudWatch alarms for these metrics on every Managed Service for Apache Flink production application. All alarms should use the Application dimension unless otherwise noted. These align with the AWS recommended alarms (opens in a new tab).

Alarm configuration notes:

• Use “3 out of 3 consecutive periods” evaluation for CPU, heap, and backpressure alarms to avoid false alarms from transient spikes during checkpoints or GC.
• For downtime, use “1 out of 1” — any downtime is immediately actionable.
• For lastCheckpointDuration, calculate the threshold from your configured checkpoint interval. If the interval is 60 seconds, alarm at > 30,000 ms.
• For millisBehindLatest / records_lag_max, combine the threshold with a trend check: a one-time spike that recovers is less concerning than a steadily increasing value.

Using Metric Trends to Drive Scaling Decisions

Point-in-time metric values can be misleading. Use trends over 30–60 minute windows to make scaling decisions:

Scale up (add KPUs) when:

• cpuUtilization Maximum is sustained above 70% for 30+ minutes during normal (non-peak) traffic. This leaves insufficient headroom for traffic spikes and checkpoint overhead.
• backPressuredTimeMsPerSecond is sustained above 100 ms/s for 15+ minutes. This means at least one operator chain cannot keep up with its input rate.
• millisBehindLatest or records_lag_max is monotonically increasing over 30+ minutes. The application is falling further behind and will not recover without more resources.
• lastCheckpointDuration is trending upward and approaching the checkpoint interval. State is growing faster than the application can snapshot it.

Scale down (reduce KPUs) when:

• cpuUtilization Maximum is sustained below 30% for 60+ minutes during peak traffic. The application is over-provisioned.
• backPressuredTimeMsPerSecond is consistently 0 for 60+ minutes. No operators are bottlenecked.
• heapMemoryUtilization Maximum is sustained below 40% for 60+ minutes. Memory is significantly over-provisioned.

Do not scale based on:

• Short spikes (< 10 minutes) in CPU or backpressure — these are often caused by checkpoint processing or transient traffic bursts and resolve on their own.
• A single high lastCheckpointDuration value — one slow checkpoint can be caused by a GC pause or S3 latency spike. Look for a trend of increasing duration across multiple checkpoints.
• numRecordsInPerSecond alone — throughput fluctuations are normal. Combine with backpressure and lag metrics to determine if the application is actually struggling.

Custom Metrics Emission

Registering Custom Metrics via Flink MetricGroup API

Flink provides four metric types through the MetricGroup API, available in any RichFunction or ProcessFunction via getRuntimeContext().getMetricGroup():

Critical: The kinesisanalytics Metric Group Requirement

Managed Service for Apache Flink publishes only metrics registered under the kinesisanalytics metric group to CloudWatch. If you register custom metrics directly on getRuntimeContext().getMetricGroup() without adding the kinesisanalytics group, your metrics will be visible in the Flink Web UI but will not appear in CloudWatch.

You must use .addGroup("kinesisanalytics") when registering custom metrics:

The kinesisanalytics group name is a legacy from when the service was called Kinesis Data Analytics. Additional metric groups can be included and will be published to CloudWatch as dimensions.

How Custom Metrics Publish to CloudWatch from Managed Service for Apache Flink

Managed Service for Apache Flink automatically publishes all Flink metrics registered under the kinesisanalytics group to CloudWatch under the AWS/KinesisAnalytics namespace. Custom metrics appear in CloudWatch with the same dimensions as built-in metrics (Application, Task, Operator, Parallelism).

Naming conventions:

• Use descriptive concatenated names: ReceivedRecordsTotal, FilteredRecordsAverage (as shown in the AWS examples)
• Prefix with a domain to avoid collisions with Flink internal metrics
• Keep names short — they contribute to CloudWatch metric cardinality and appear in dashboards

Code Example

The following example demonstrates the pattern from the official AWS CustomMetrics sample (opens in a new tab). It uses a RichMapFunction as a pass-through that emits a counter and a gauge:

This mapper is then used in the pipeline to emit metrics at different stages:

Cardinality Limits and Management Strategies

CloudWatch has practical limits on custom metric cardinality that affect cost and usability:

• CloudWatch pricing: Each unique metric (combination of namespace, metric name, and dimension values) is billed as a custom metric. At $0.30 per metric per month, an application with 100 custom metrics × 32 subtasks = 3,200 billable metrics = ~$960/month.
• Dashboard limits: CloudWatch dashboards become slow and unwieldy with more than a few hundred metrics. Operator-level custom metrics at high parallelism quickly exceed this.
• API throttling: CloudWatch PutMetricData has API rate limits. Excessive custom metrics can cause metric publishing delays or dropped data points.

Strategies for managing cardinality:

1. Register metrics at the operator level, not per-key. Do not create a separate counter for each key value (e.g., orders.processed.customerA, orders.processed.customerB). Instead, use a single counter per operator and rely on subtask-level dimensions for distribution analysis.

2. Use Application-level aggregation for alarms. CloudWatch automatically aggregates subtask-level metrics to the Application level using the Average, Sum, Min, and Max statistics. Set alarms on Application-level aggregates rather than individual subtask metrics.

3. Limit custom metrics to 10–20 per operator. Beyond this, the cardinality cost grows quickly with parallelism. Focus on metrics that directly inform operational decisions (error rates, processing latency, business KPIs).

4. Prefer counters and gauges over histograms. Histograms publish multiple CloudWatch metrics per registration (min, max, mean, p50, p75, p95, p99, count) — a single histogram registration produces 8+ CloudWatch metrics per subtask.

Metric Reporting Interval and Cost Implications

Managed Service for Apache Flink reports metrics to CloudWatch at a fixed interval, typically every 60 seconds (with 4 snapshots per minute for throughput metrics). This has several implications:

• Granularity: Custom metrics have 1-minute resolution in CloudWatch. Sub-minute fluctuations are averaged within each reporting interval. For latency-sensitive monitoring, use the Flink Web UI (which updates in near-real-time) alongside CloudWatch.
• Cost: CloudWatch charges are per-metric per-month, not per-data-point. The reporting interval does not directly affect cost — but more frequent reporting (if configurable in future Managed Service for Apache Flink versions) would not increase cost for the same metric set.
• Meter accuracy: Flink’s MeterView calculates rates over a configurable window (e.g., 60 seconds). Align the meter window with the reporting interval for consistent rate values in CloudWatch. A 60-second MeterView window with 60-second reporting produces stable rate metrics.
• Gauge staleness: Gauges report their current value at each reporting interval. If the gauge value changes rapidly between intervals, CloudWatch only captures the value at the reporting moment. For rapidly changing values, consider using a counter (cumulative) and computing rates in CloudWatch metric math.

Monitoring Setup Operations

Operational specifics for wiring up monitoring on a new or existing MSF application.

Log Group / Stream Creation Order

The log group and stream must exist before add-application-cloud-watch-logging-option is called — MSF does not create them. Application logs are silently dropped if the destination doesn’t exist.

Default CloudWatch log retention is unlimited — set retention explicitly to control cost. 30 days is a reasonable default; 7 days for dev.

MetricsLevel — Cost vs Granularity

OPERATOR/TASK/PARALLELISM are useful for short-window diagnosis; switch back to APPLICATION after. CloudWatch custom metrics are billed at ~$0.30/metric-month.

treat-missing-data Semantics

The --treat-missing-data flag determines alarm state when a metric stops reporting. The right value differs per alarm:

Cumulative counters need special treatment. numberOfFailedCheckpoints, fullRestarts, and numRestarts only ever increase during a job’s lifetime; they reset only when the application restarts. Three rules follow from this and must be configured together — getting any one wrong breaks the alarm.

1. Alarm on RATE(), not the raw value. RATE(numberOfFailedCheckpoints) > 0 alarms when checkpoints are actively failing in the current window. An alarm on the raw cumulative value would trip on the first failure ever and stay in ALARM until the next application restart, even after the underlying issue was fixed, because the count never decreases.

2. Set treat-missing-data: notBreaching. Some Flink versions only emit these counters when the value is non-zero, so a healthy app produces emission gaps as a normal steady state. notBreaching makes those gaps count as “not in alarm” instead of flipping the alarm to INSUFFICIENT_DATA. Combined with rule 1, the lifecycle is: healthy app → no data points → OK; failures land → non-zero RATE() → ALARM; failures stop → gaps return → OK.

3. Route the alarm as an investigation trigger, not a self-clearing health signal. Rule 1 means the alarm will return to OK on its own once new failures stop landing in the window. That is a statement about recent activity, not about resolution: “no new failures in the last N minutes” is not the same as “an operator diagnosed the root cause.” Send these alarms to a ticketing or acknowledgment workflow where a human closes them after investigation, not to a paging path that treats the OK transition as “problem solved.” A burst of checkpoint failures that stops on its own is still worth a postmortem.

Alarm Priority Order for New Applications

When setting up monitoring on a new app, configure in this order — earlier alarms catch issues that downstream alarms can’t:

1. RATE(numberOfFailedCheckpoints) > 0 — checkpoint failures = data loss risk
2. downtime > 0 (Flink 1.x) or restartingTime > 0 (Flink 2.2) — app not processing
3. RATE(fullRestarts) > 0 (Flink 1.x) or RATE(numRestarts) > 0 (Flink 2.2) — instability
4. cpuUtilization > 80% Maximum, 3/3 — capacity warning
5. heapMemoryUtilization > 90% Maximum, 3/3 — OOM risk (AWS’s officially recommended alarm). Substitute or supplement with containerMemoryUtilization > 85% for stateful apps where off-heap RocksDB / native memory matters; AWS’s metric docs call container “a better tracker of total memory exhaustion.”
6. backPressuredTimeMsPerSecond > 500 Average, 3/3 — bottleneck
7. millisBehindLatest (Kinesis) or records_lag_max (Kafka) — falling behind input

For exact thresholds and statistic recommendations, see Recommended Alarms and Thresholds above.

Common Setup Mistakes