Netflix currently runs two different autoscalers for Apache
Flink, and they are actively moving toward replacing the
custom one with the open-source Kubernetes-native
autoscaler.
The custom autoscaler predates the Kubernetes operator by several years. It was built when the platform ran Flink on its own scheduling layer, and it carries assumptions from that era — most visibly that scaling decisions are made from a single aggregated backlog metric rather than per-vertex.
Feature parity is the gating question. The open-source autoscaler reached rough equivalence on throughput-based scaling last year, but two things kept the custom one alive: its handling of sources with irregular partition counts, and the operational tooling built around it.
What follows is a comparison of how each one reacts to the same three production incidents, and what that implies for the migration order.
The custom autoscaler predates the Kubernetes operator by several years. It was built when the platform ran Flink on its own scheduling layer, and it carries assumptions from that era — most visibly that scaling decisions are made from a single aggregated backlog metric rather than per-vertex.
Feature parity is the gating question. The open-source autoscaler reached rough equivalence on throughput-based scaling last year, but two things kept the custom one alive: its handling of sources with irregular partition counts, and the operational tooling built around it.
What follows is a comparison of how each one reacts to the same three production incidents, and what that implies for the migration order.
Summary
Netflix runs two Flink autoscalers today and is consolidating
on the open-source, Kubernetes-native one — the custom
version is being retired once feature parity closes the
remaining gap.
01 / 06 · Reading
Read it, ask what's in it, or have it read to you
The reader strips the page down to its text. The summary button sends that text to Gemini Nano, on the phone. Nothing is uploaded.
The waveform button is new: the phone reads the article aloud itself, no separate sync step required. It keeps playing behind whatever screen you open next, in the same floating bar Readback already used.