Damn that's some scope creep if I ever saw it: 'try sending Arrow frames end to end' => 'rewrite the otel pipeline in rust'. Seems like the goals of the contributors don't exactly align with the goals of the project.
Kind of a bummer - one thing i was hoping to come out of this was better Arrow ecosystem support for golang.
I've just started exploring adding OpenTelemetry support to the Comet subproject of DataFusion. I'm excited to see the integration with Apache Arrow (Rust) and potentially DataFusion in the future.
Thanks for sharing this — it’s a really promising direction. The advantages of Arrow for OTLP, especially when used end-to-end, are compelling given the protocol overhead of OTLP.
We’ve been thinking along similar lines with the use of Rust, particularly for OpenTelemetry collection in environments where high performance and low resource overhead are critical, such as edge and serverless. With that in mind, we’ve open-sourced a lightweight OpenTelemetry collector written in Rust to address these use cases. We’ve also developed a native Lambda extension around it, and have seen encouraging interest from folks aiming to improve cold start times.
The project is still fairly early, but we’re optimistic that Rust can open up new opportunities for efficient observability pipelines. Vendors like Datadog are also moving in this direction with their Lambda extension and appear to be adopting Rust more broadly for data-plane components.
If this resonates, feel free to take a look here: https://github.com/streamfold/rotel. We’d love to hear your thoughts on how this could be useful.
Integrations with OTLP are critical to driving adoption and probably one of the biggest pain points we've encountered when adopting it ourselves (and encouraging others to the same).
Adopting OTLP without third-party support is pretty time consuming, especially is your tech stack is large and/or varied.
Re runtimes: curious about this too. Feels like the right direction if you’re optimizing a telemetry pipeline.
Performance optimization and being able to "plug in" to the data ecosystem that Apache Arrow exists in.
OpenTelemetry is pretty great for a lot of uses, but the protocol over the wire is too chunky for some applications where. From last year's post on the topic[0]:
> In a side-by-side comparison between OpenTelemetry Protocol (“OTLP”) and OpenTelemetry Protocol with Apache Arrow for similarly configured traces pipelines, we observe 30% improvement in compression. Although this study specifically focused on traces data, we have observed results for logs and metrics signals in production settings too, where OTel-Arrow users can expect 50% to 70% improvement relative to OTLP for similar pipeline configurations.
For your average set of apps and services running in a k8s cluster somewhere in the cloud, this is just a nice-to-have, but size on wire is a problem for a lot of systems out there today, and they are precluded from adopting OpenTelemetry until that's solved.
Not sure, but seems like it will be producing apache arrow data and carrying it across the data stack end to end from OTEL. This would be great for creating data without a bunch of duplication/redundant processing steps and exporting it in a form that's ready to query.
Unless I dont understand that fully (which could be the case).
This idea could fly if downstream readers will be able to read it. Json is great because anything can read it, process, transform and serialize without having to know the intrisics of the protocol.
Whats the point of using binary, columnar format for data in transit?
Arrow is an in-memory columnar format, kinda orthogonal to parquet (which is an at-rest format). Protobuf is a better comparison, but it's more message oriented and not suited for analytics.
Man Ive dipped my toes into this too, and yeah, the way everyone wants different things always shakes things up fast. Kinda love seeing where it all ends up tbh.
Damn that's some scope creep if I ever saw it: 'try sending Arrow frames end to end' => 'rewrite the otel pipeline in rust'. Seems like the goals of the contributors don't exactly align with the goals of the project.
Kind of a bummer - one thing i was hoping to come out of this was better Arrow ecosystem support for golang.
I've just started exploring adding OpenTelemetry support to the Comet subproject of DataFusion. I'm excited to see the integration with Apache Arrow (Rust) and potentially DataFusion in the future.
[dead]
Thanks for sharing this — it’s a really promising direction. The advantages of Arrow for OTLP, especially when used end-to-end, are compelling given the protocol overhead of OTLP.
We’ve been thinking along similar lines with the use of Rust, particularly for OpenTelemetry collection in environments where high performance and low resource overhead are critical, such as edge and serverless. With that in mind, we’ve open-sourced a lightweight OpenTelemetry collector written in Rust to address these use cases. We’ve also developed a native Lambda extension around it, and have seen encouraging interest from folks aiming to improve cold start times.
The project is still fairly early, but we’re optimistic that Rust can open up new opportunities for efficient observability pipelines. Vendors like Datadog are also moving in this direction with their Lambda extension and appear to be adopting Rust more broadly for data-plane components.
If this resonates, feel free to take a look here: https://github.com/streamfold/rotel. We’d love to hear your thoughts on how this could be useful.
Integrations with OTLP are critical to driving adoption and probably one of the biggest pain points we've encountered when adopting it ourselves (and encouraging others to the same).
Adopting OTLP without third-party support is pretty time consuming, especially is your tech stack is large and/or varied.
Re runtimes: curious about this too. Feels like the right direction if you’re optimizing a telemetry pipeline.
Wow, anyone able to provide a ELI5? OTel sounds amazing but this is flying over my head
Warning: this is an oversimplification.
Performance optimization and being able to "plug in" to the data ecosystem that Apache Arrow exists in.
OpenTelemetry is pretty great for a lot of uses, but the protocol over the wire is too chunky for some applications where. From last year's post on the topic[0]:
> In a side-by-side comparison between OpenTelemetry Protocol (“OTLP”) and OpenTelemetry Protocol with Apache Arrow for similarly configured traces pipelines, we observe 30% improvement in compression. Although this study specifically focused on traces data, we have observed results for logs and metrics signals in production settings too, where OTel-Arrow users can expect 50% to 70% improvement relative to OTLP for similar pipeline configurations.
For your average set of apps and services running in a k8s cluster somewhere in the cloud, this is just a nice-to-have, but size on wire is a problem for a lot of systems out there today, and they are precluded from adopting OpenTelemetry until that's solved.
[0]: https://opentelemetry.io/blog/2024/otel-arrow-production/
Not sure, but seems like it will be producing apache arrow data and carrying it across the data stack end to end from OTEL. This would be great for creating data without a bunch of duplication/redundant processing steps and exporting it in a form that's ready to query.
Unless I dont understand that fully (which could be the case).
This idea could fly if downstream readers will be able to read it. Json is great because anything can read it, process, transform and serialize without having to know the intrisics of the protocol.
Whats the point of using binary, columnar format for data in transit?
better compression https://opentelemetry.io/blog/2023/otel-arrow/
You don't do high performance without knowing the data schema.
Is Arrow better than Parquet or Protobuf?
Arrow is an in-memory columnar format, kinda orthogonal to parquet (which is an at-rest format). Protobuf is a better comparison, but it's more message oriented and not suited for analytics.
Not having to write to disk is great, and zero-copy in memory access is instant...
the blog post comparison is against OTLP which is protobuf
[dead]
This diagram really depicts things nicely
https://opentelemetry.io/blog/2023/otel-arrow/row-vs-columna...
A bit hand wavy.
> We are interested in making OTAP pipelines safely embeddable, through strict controls on memory and through support for thread-per-core runtimes.
I'm curious about the thread-per-core runtimes, are there even any mature thread-per-core runtimes in Rust around?
glommio is pretty well respected. https://www.datadoghq.com/blog/engineering/introducing-glomm... https://github.com/DataDog/glommio
ByteDance also has their very fast monio. https://github.com/bytedance/monoio
Both integrate io-uring support for very fast io.
[dead]
Man Ive dipped my toes into this too, and yeah, the way everyone wants different things always shakes things up fast. Kinda love seeing where it all ends up tbh.
[dead]