/developing-applications-on-managed-service-for-apache-flink

MSF Pricing Calculator

Overview

Estimate MSF application monthly cost for sizing decisions, budgeting, and optimization analysis. All prices below are us-east-1 — verify against MSF pricing page for other regions.

Pricing Components

There is no separate charge for application code in S3 (your S3 bucket cost only) and no per-API-call charge for Flink REST API or kinesisanalyticsv2 operations.

Formula

KPU_count   = Parallelism / ParallelismPerKPU
Total_KPUs  = KPU_count + 1                                    # orchestration

KPU_cost     = Total_KPUs × $0.11 × hours_per_month
storage_cost = max(0, state_GB - KPU_count × 50) × $0.10
backup_cost  = total_snapshot_GB × $0.023

monthly_total = KPU_cost + storage_cost + backup_cost

24/7 = 730 hours/month. Most production apps stay under the 50 GB/KPU storage included tier.

Sizing → Cost Reference (24/7, $0 storage)

These assume stateless-to-moderate stateful workloads. Stateful workloads (keyed windows with large state) typically require PPK=1, doubling KPU count and cost vs. the stateless equivalent.

Worked Examples

Small / development (Parallelism=2, PPK=1, 24/7, negligible state, no snapshots):

• KPU count: 2 / 1 = 2 application KPUs, plus 1 orchestration KPU = 3 billed KPUs
• KPU cost: 3 × $0.11 × 730 = $240.90
• Running application storage: state ≤ 2 × 50 GB included → $0 (the 50 GB/KPU allocation is included, not separately billable; do not multiply 50 × $0.10 per KPU)
• Snapshot cost: $0 (none retained)
• Total: ~$241/month

The orchestration KPU is the most common sizing mistake at this scale: it raises the bill from $158 (2 KPUs) to $241 (3 KPUs), a 52% increase. Always include it.

Medium production (8 parallel, PPK=1, 24/7, 5 GB state):

• KPUs: 8 + 1 = 9 → 9 × $0.11 × 730 = $722.70
• Storage: 5 GB ≤ 8 × 50 GB included → $0
• Snapshots: 5 GB × $0.023 = $0.12
• Total: ~$723/month

Enterprise with autoscaling (avg 16 KPU, peak 32 KPU 20% of time, 60 GB state):

• Peak: 33 × $0.11 × 146 = $530
• Off-peak: 17 × $0.11 × 584 = $1,092
• Snapshots: 60 GB × $0.023 = $1.38
• Total: ~$1,623/month

Cost Optimization Levers

Ranked by typical impact:

1. Right-size PPK for workload: Stateless apps at PPK=2 cost half of PPK=1 at the same parallelism. Rule of thumb: stateless → PPK=2; stateful → PPK=1.
2. Stop non-prod when idle: Dev/test apps run 200h/month vs 730h is a 73% reduction. Snapshot before stop, restore on start.
3. Enable autoscaling for variable load: 30–50% average reduction for spiky workloads with predictable off-peaks.
4. Match Parallelism to source partitions: Idle subtasks waste KPUs. Parallelism > shard/partition count creates them.
5. Snapshot retention strategy: Without retention, a streaming app's snapshot footprint grows monotonically.
6. Always remember the orchestration KPU: At 1–2 KPU sizing it's 33–50% of the bill — easy to under-estimate.

Common Mistakes

References

• MSF Pricing
• resource-optimization.md — KPU sizing inputs for the calculator
• scaling-decisions.md — cost impact of scaling actions