These benchmarks show typical upper-bound times for Savant on structured, tabular data under the test conditions below. Actual times vary with data shape, workflow design, and customer network and source-system conditions.
Test conditions
Data. Benchmark tables have four columns: integer, text, number and date.
Runs. Each figure is the average of three runs on Savant's standard managed compute.
Scope. Benchmarks cover structured, tabular data (database tables and delimited, Excel or Parquet files). They don't cover unstructured inputs such as PDFs, images or documents, or workflow steps that call an AI model while the workflow runs. Workflows built with Savvy are covered – the exclusion applies only to AI steps inside a run.
Task agent performance
Task agents are the data processing steps that make up a Savant workflow. Times cover processing within Savant only, excluding reading from or writing to external systems. The Filter, Pivot, Unpivot and Stack task agents use standard configurations. Transform uses typical text, date, math, aggregation and window formulas. Blend uses an inner join on two key columns.
Task agent | 100K rows | 1M rows | 10M rows | 100M rows |
Pivot | 5 sec | 15 sec | 30 sec | 3 min |
Unpivot | 5 sec | 15 sec | 25 sec | 45 sec |
Filter | 5 sec | 15 sec | 30 sec | 3 min |
Stack | 5 sec | 10 sec | 10 sec | 90 sec |
Transform | 5 sec | 15 sec | 45 sec | 6 min |
Blend | 15 sec | 45 sec | 90 sec | 5 min |
Connector performance
Times are measured with the source or destination in the same cloud region as the Savant workspace, over a standard connection. Customer network policies (firewalls, private connectivity, proxies), source-system load and API rate limits can lengthen them.
Database read. SQL databases such as Snowflake, Databricks and SQL Server, with similar results expected across common SQL databases: one 100M-row table in under 30 minutes; one 200M-row table in under 45 minutes.
Database write. One 100M-row table in under 3 hours; one 200M-row table in under 4.5 hours.
File system read. Cloud file storage (benchmarked on OneDrive, Google Drive and Box): 100 CSV files of 1M rows each in under 40 minutes; 100 Excel files of 1M rows each in under 4 hours.
File system write. 100M rows of CSV output (100 files of 1M rows, or 1,000 of 100K) in under 4 hours; the same volume as Excel in under 6 hours.
Cloud storage read. Object storage (such as Amazon S3, Azure Blob Storage and Google Cloud Storage): one 50M-row CSV in under 30 minutes.
Cloud storage write. One 50M-row CSV in under 2 hours.
Actual read and write times may be significantly shorter with Savant's intelligent caching, which optimizes data movement and avoids redundant transfers.
Concurrent processing
Elastic compute. Savant runs workflows on managed cloud compute that scales automatically with workload – no infrastructure to provision or tune.
No queueing across workflows. Different workflows run concurrently without waiting on one another, so customers can schedule as many concurrent workflows as they need, subject to the limits below.
One run per workflow at a time. To prevent data collisions, each workflow runs one instance at a time. Workflows that power analytic apps serving multiple business users are designed to support up to three concurrent runs.
Limits. Scaling is subject to the customer's subscription and the underlying cloud provider's capacity and service quotas.
