Adtechs use the aggregation service to generate summary reports from aggregatable reports (see the workflow). The aggregation service can process jobs of different input sizes (reports and pre-declared aggregation buckets). Each report can include a different number of events. Each event is a key-value pair in an aggregatable report's encrypted payload. A pre-declared aggregation bucket is an entry (also referred to as domain key in this document) in the pre-declared buckets file that is provided as an input in the aggregation job (see createJob API documentation). Reports used in generating this guidance each have a payload with 10 events and were padded to the report size of 20 (standard for Attribution Reporting API and Shared Storage). Reports and domain keys are sharded into 1000 files individually when running the benchmark.

The table below provides cloud instance type guidelines for minimum required instance memory for aggregation jobs with different sizes of reports and domain keys. For example, if your individual jobs are around the 10K reports/50 million domain keys, an instance type with at least 32 GB memory is required.

Note that this is a general guideline based on the input data as mentioned above and adtechs may use a different instance type depending on their specific needs. For example, if an adtech wants lower latency results, one may need to provision larger compute capacity instance types

After the instance type has been selected, the running times provided below can be used to estimate the cost per job and the total cost:

• Cost per job = job running time * cost of selected instance type (AWS pricing, GCP pricing)
• Total estimated cost = # of jobs * cost per job

Adtechs may use the formula above to determine the most cost-efficient machine for their jobs. For example, an adTech chooses a r5.12xlarge instance for jobs with 1 million aggregatable reports and 200 million output domain keys. The running time for a job of this size is 1 hour per job:

• Instance cost: $3.024 / hour
• Cost per job = 1 hour * $3.024 / hour = $3.024
• Total cost = 10 jobs * $3.024 / job = $30.24

The table below considers the following AWS and GCP instances types:

• AWS: m5.2xlarge (default), m5.4xlarge, m5.8xlarge, m5.12xlarge, r5.2xlarge, r5.4xlarge, r5.8xlarge, r5.12xlarge
• GCP: n2d-standard-8 (default), n2d-standard-16, n2d-standard-32, n2d-standard-48, n2d-highmem-8, n2d-highmem-16, n2d-highmem-32, n2d-highmem-48, n2d-highmem-64, n2d-highmem-80, n2d-highmem-96

On AWS, if your jobs have larger domain keys (~1 billion), you may use the AWS machine type x2iedn.8xlarge. For any other questions, please feel free to reach out to Aggregation Support.

For a given aggregation job, the rows (R) correspond to the input reports, the columns (D) correspond to the domain keys (in millions), and the value is the minimum workable instance with memory (GiB).

The aggregation job completion time is the time an instance picks up the job from the job queue to when the results were written to storage. The job completion time can be calculated from the fields returned from the getJob API using request_updated_at minus request_processing_started_at field. It depends on a number of factors such as the size of the job inputs, the service instance compute capacity, the auto-scaling size, and more.

The aggregation job completion times are split into different brackets. For example, the "00 hr 40 min" bracket indicates that the job ran for more than 20 minutes but less than or equal to 40 minutes. The "n/a" bracket indicates that the instance was not able to process the job, and a higher capacity instance is needed.

The processing time values here are subject to change.

In the following table, the rows (R) correspond to the input reports. The columns (D) correspond to the domain keys (in millions).

• Memory(GiB): 32 / vCPU: 8

• Memory(GiB): 64 / vCPU: 8

• Memory(GiB):64 / vCPU: 16

• Memory(GiB): 128 / vCPU: 16

• Memory(GiB): 128 / vCPU: 32

• Memory(GiB): 256 / vCPU: 32

• Memory(GiB): 256 / vCPU: 48

• Memory(GiB): 384 / vCPU: 48

• Memory(GiB): 512 / vCPU: 64

• Memory(GiB): 640 / vCPU: 80

• Memory(GiB): 768 / vCPU: 96