To optimize Dataproc Serverless and Dataflow Prime costs, work the Data Compute Unit, the billing unit both share. Measure DCU-hours per job, right-size workers so you burn fewer DCUs per vCPU and GB, cut the data shuffled, run fault-tolerant work on Spot, and tune autoscaling so jobs release resources promptly. Dataproc Serverless is about 0.06 US dollars per DCU-hour with no idle cost. Because both services charge for compute consumed rather than provisioned, every DCU you avoid is a direct saving.
Last updated: June 2026. Written by Fredrik Filipsson and reviewed by Morten Andersen, built on our See, Cut, Lock, Run method.
This article sits in our complete guide to Google Cloud cost optimization, the cluster pillar it links up to, and is part of the broader cloud cost optimization playbook. Trimming data-processing compute is a Cut step move. For the analytics side of the GCP data stack, see our companion guide on reducing BigQuery slot and reservation costs.
Both services bill the DCU, so cut DCUs, not nodes. A Data Compute Unit measures the compute a job consumes, so the work is the same on Dataproc Serverless and Dataflow Prime: shrink workers, move less data, and release resources fast. The expensive mistake is over-provisioning memory per vCPU or shuffling more data than the job needs, both of which inflate the DCU count directly.
What is a Data Compute Unit?
A Data Compute Unit, or DCU, is the billing unit shared by Dataproc Serverless and Dataflow Prime. One DCU is comparable to the resources used by a worker with 1 vCPU and 4 GB of memory running for one hour. In Dataproc Serverless the calculation is more granular: each vCPU counts as 0.6 DCU, each GB of RAM up to 8 GB per vCPU adds 0.1 DCU, and each GB above 8 GB per vCPU adds 0.2 DCU, per the Dataproc Serverless pricing documentation. Dataflow Prime prices the same way, with DCUs representing the compute resources allocated to run your pipeline, per the Dataflow pricing documentation. The lesson from the formula is that fat memory per vCPU is expensive: memory above the 8 GB threshold costs twice as much DCU per GB.
How much does Dataproc Serverless cost?
Dataproc Serverless for Apache Spark is about 0.06 US dollars per DCU-hour with no idle cost, plus shuffle storage and any accelerators. DCUs and shuffle storage are billed per second with a one-minute minimum charge, while accelerators carry a five-minute minimum, per the Dataproc Serverless pricing documentation. Because there is no idle cost, a batch that runs for ten minutes is billed for roughly ten minutes of DCUs, which is what makes serverless attractive for intermittent Spark work. Dataflow Prime DCU rates vary by region and pipeline type; confirm the current per-DCU figures for both services in the linked documentation before modeling, because Google Cloud data-processing prices change.
How do you cut DCU consumption?
Cut DCU consumption by right-sizing workers, reducing the data shuffled, and using cheaper capacity. The steps below apply to both Dataproc Serverless and Dataflow Prime.
- Measure DCU consumption. Read the DCU-hours per job from the billing export so you know which pipelines and Spark batches drive the cost.
- Right-size workers. Reduce vCPU and memory per worker to the minimum the job needs, since each vCPU is 0.6 DCU and memory above 8 GB per vCPU costs 0.2 DCU per GB instead of 0.1.
- Cut the data shuffled. Filter early, use combiners and efficient coders, and avoid wide shuffles so less data moves and fewer DCU-seconds and shuffle storage are billed.
- Use Spot and the right tier. Run fault-tolerant batches on Spot or preemptible capacity and choose the standard tier unless a premium tier feature is genuinely required.
- Tune autoscaling and idle timeouts. Set autoscaling bounds and idle timeouts so jobs release workers promptly and never hold DCUs after the work is done.
When is serverless cheaper than a cluster?
Serverless is cheaper for spiky or intermittent work; a steady cluster is cheaper for constant load. Dataproc Serverless has no idle cost and bills only for the DCUs a job consumes, so for batch jobs that run a few times a day or on irregular schedules it avoids paying for a cluster that sits waiting. A long-running Dataproc cluster on committed or Spot VMs, by contrast, can be cheaper when work runs continuously and keeps the cluster busy, because the committed and Spot rates on steady utilization beat per-DCU serverless pricing. The decision is how continuously the work runs, not the service name. The full data-processing playbook lives in our Google Cloud guide.
Burning more DCUs than your pipelines need?
Our Google Cloud cost audit profiles your Dataproc and Dataflow jobs by DCU, right-sizes workers, finds the wide shuffles, and moves fault-tolerant work to Spot. On the performance model, you pay only from realized savings. No savings, no fee.
Book a Google Cloud cost audit →What inflates a Dataproc or Dataflow bill?
The biggest data-processing cost mistakes are oversized workers, wide shuffles, on-demand capacity for fault-tolerant work, and jobs that hold resources after they finish. Memory set well above 8 GB per vCPU quietly doubles the DCU cost of that memory, so default-fat workers are a common silent overspend. Wide shuffles move far more data than necessary and bill both DCU-seconds and shuffle storage. Running batch jobs on on-demand instead of Spot leaves a discount on the table. And pipelines that do not release workers promptly keep billing DCUs for idle capacity. Working the five steps above removes each of these in turn.
The Google Cloud Cost Optimization Field Guide includes the DCU right-sizing worksheet and the shuffle-reduction checklist referenced above. It is the downloadable companion to this guide.
Frequently asked questions
What is a Data Compute Unit?
A Data Compute Unit, or DCU, is the billing unit for Dataproc Serverless and Dataflow Prime. One DCU is comparable to the resources of a worker with 1 vCPU and 4 GB of memory running for one hour. In Dataproc Serverless, each vCPU counts as 0.6 DCU and memory adds 0.1 DCU per GB up to 8 GB per vCPU, then 0.2 DCU per GB above that.
How much does Dataproc Serverless cost?
Dataproc Serverless for Apache Spark is about 0.06 US dollars per DCU-hour with no idle cost, plus shuffle storage and any accelerators. DCUs and shuffle storage are billed per second with a one-minute minimum, and accelerators have a five-minute minimum. Confirm the current rate in the Dataproc Serverless pricing documentation.
How do I reduce Dataflow Prime DCU usage?
Reduce Dataflow Prime DCUs by cutting the data your pipeline shuffles, right-sizing workers, using efficient coders and combiners, and letting vertical and horizontal autoscaling release resources. Since DCUs represent the compute allocated to the pipeline, less work and tighter resource allocation directly lower the DCU count and the bill.
Is Dataproc Serverless cheaper than a Dataproc cluster?
Dataproc Serverless is usually cheaper for spiky or intermittent Spark work because it has no idle cost and bills only for the DCUs a job consumes. A long-running, steadily busy Dataproc cluster on committed or Spot VMs can be cheaper for constant load. Match the model to how continuously the work runs.
The short version
Dataproc Serverless and Dataflow Prime both bill the DCU, so cut DCUs: right-size workers, watch the 8 GB-per-vCPU memory threshold, shuffle less, run on Spot, and release resources fast. Serverless wins on intermittent work, a steady cluster on constant load. When you want your pipelines profiled and tuned, that is what our Google Cloud cost optimization service delivers.
Cloud pricing and service behavior change frequently. Verify the specifics in this guide against the providers’ own current documentation and the FinOps Foundation: Google Cloud pricing ↗, Google Cloud documentation ↗ and FinOps Foundation Framework ↗. This article also reflects Cloud Cost Room’s hands-on, vendor-neutral engagement experience.