Azure Functions Trace 2019

This is a set of files representing part of the workload of Microsoft's Azure Functions offering, collected in July of 2019. This dataset is a subset of the data described in, and analyzed, in the USENIX ATC 2020 paper 'Serverless in the Wild: Characterizing and Optimizing the Serverless Workload at a Large Cloud Provider'.

Functions in Azure Functions are grouped into Applications. Included here is only data pertaining to a random sample of Azure Functions applications. The sampling is done per application, so that if there is data about an application in the trace, then all of its functions are included. The sampling rate is small and unspecified, but as the accompanying notebook shows, the distributions in the released trace are a good match to those in the ATC paper.

In Azure Functions, applications are the unit of resource allocation. This has a few practical implications: for example, warm-up decisions are made at the application level, and memory allocation is measured by application, not by function. The 'HashOwner' field in these files is used to group applications that belong to the same subscription in Azure. It is included to indicate applications that are possibly related to each other.

The dataset comprises this description, and an R notebook with plots comparing the released trace with the ATC paper, and the following sets of files:

Function invocation counts and triggers
Function execution time distributions
Application memory allocation distributions