openmeter VS metering-operator

This plugin guarantees exactly-once delivery between Kafka topics and ClickHouse tables, which is critical, as Kafka Connect tasks are only aware of the latest topic offset acknowledged by the consumer. For example, consumers can fail to acknowledge a processed offset due to a network error or an exception. This is great as exactly-once inserts prevent dropping or double-inserting usage, leading to incorrect billing.

In OpenMeter, we pre-aggregate usage events into one-minute tumbling windows to reduce the number of rows we need to scan at query time. To do this, with ClickHouse, we use the AggregatingMergeTree table engine that enables incremental data aggregation when combined with MaterializedView. In ClickHouse, MaterializedViews are trigger-based and update when new records are inserted into the source table. Consequently, the corresponding materialized views are updated whenever Kafka Connect transfers a batch of events to ClickHouse. This also means inserts can fail when the view cannot process a record at trigger. We send failed events into the Dead Letter Queue topic for later processing.

To help ClickHouse with hot topics, we will consider adding an extra streaming aggregation step for high-producers, but this time with a more horizontally scalable stream processor like Arroyo. This would reduce ClickHouse insert batch sizes. Based on our tests, ClickHouse works best if batch sizes are 50-100k and less frequent than per second.

To see it in action, check out our open-source repo: https://github.com/openmeterio/openmeter

I’m researching how to meter a POD execution duration for auditable use cases like billing. We are working on an infrastructure metering solution, and our current example polls the k8s API every second to see which pods are running: see code here - wondering what folks think about this approach.

OpenMeter leverages stream processing to update meters in real-time while simultaneously processing large volumes of events. The core is written in Go and uses the CloudEvents format to describe usage, Kafka to ingest events, and ksqlDB to dedupe and aggregate meters. We are also working on a Postgres sink for long-term storage. Check out our GitHub to learn more: https://github.com/openmeterio/openmeter

Metering Operator. But looks like it's not under active development.