Improved concurrency model
New parallel processing with improved performance
We’ve released an improved concurrency model that provides measurable performance improvements and enables more efficient vertical scaling.
Previously, Vector was limited to executing transforms on a single thread per transform, which led to them often being the bottleneck in Vector pipelines.
With this new release, Vector will perform faster when a transform is a bottleneck, assuming that more CPUs are available to share their work. This works by Vector determining whether an individual transform will be processed in parallel when there is a sufficient load to the environment.
This improvement works by spinning up multiple short-lived tasks that concurrently run the same transform logic on separate batches of events. No configuration changes are necessary to start taking advantage of this feature. The current list of transforms that support parallelization are:
metric_to_log
filter
log_to_metric
remap
But we will be rolling this out this support to more transforms soon.
If you any feedback for us, let us know on our Discord chat or on Twitter.