Blanket statements like this miss the point. Not all data is waste. Especially high-cardinality, non-sampled traces. On a 4-core ClickHouse node, we handled millions of spans per minute. Even short retention windows provided critical visibility for debugging and analysis.

Sure, we should cut waste, but compression exists for a reason. Dropping valuable observability data to save space is usually shortsighted.

And storage isn't the bottleneck it used to be. Tiered storage with S3 or similar backends is cheap and lets you keep full-fidelity data without breaking the budget.

▲

ofrzeta a day ago | parent | next [-]

> Dropping valuable observability data to save space is usually shortsighted

That's a bit of a blanket statement, too :) I've seen many systems where a lot of stuff is logged without much thought. "Connection to database successful" - does this need to be logged on every connection request? Log level info, warning, debug? Codebases are full of this.

	▲	nijave a day ago \| parent \| next [-]
		Yes, it allows you to bisect a program to see the block of code between log statements where the program malfunctioned. More log statements slice the code into smaller blocks meaning less places to look.
	▲	citrin_ru a day ago \| parent \| prev \| next [-]
		Probably not very useful for prod (non debug) logging, but it’s useful when such events are tracked in metrics (success/failure, connect/response times). And modern databases (including ClickHouse) can compress metrics efficiently so not much space will be spent on a few metrics.
	▲	vidro3 19 hours ago \| parent \| prev \| next [-]
		in our app each user polls for a resource availability every 5 mins. do we really need "connection successful" 500x per minute? i dont see this as breaking up the logs into smaller sections. i see it as noise. i'd much rather have a ton of "connection failed" whenever that occurs than the "success" constantly
	▲	throwaway0665 a day ago \| parent \| prev [-]
		There's always another log that could have been key to getting to the bottom of an incident. It's impossible to know completely what will be useful in advance.

▲

jiggawatts a day ago | parent | prev [-]

I agree with both you and the person you're replying to, but...

My centrist take is that data can be represented wastefully, which is often ignored.

Most "wide" log formats are implemented... naively. Literally just JSON REST APIs or the equivalent.

Years ago I did some experiments where I captured every single metric Windows Server emits every second.

That's about 15K metrics, down to dozens of metrics per process, per disk, per everything!

There is a poorly documented API for grabbing everything ('*') as a binary blob of a bunch of 64-bit counters. My trick was that I then kept the previous such blob and simply took the binary difference. This set most values to zero, so then a trivial run length encoding (RLE) reduced a few hundred KB to a few hundred bytes. Collect an hour of that, compress, and you can store per-second metrics collected over a month for thousands of servers in a few terabytes. Then you can apply a simple "transpose" transformation to turn this into a bunch of columns and get 1000:1 compression ratios. The data just... crunches down into gigabytes that can be queried and graphed in real time.

I've experimented with Open Telemetry, and its flagrantly wasteful data representations make me depressed.

Why must everything be JSON!?

▲

nijave a day ago | parent | next [-]

I think Prometheus works similar to this with some other tricks like compressing metric names.

OTEL can do gRPC and a storage backend can encode that however it wants. However, I do agree it doesn't seem like efficiency was at the forefront when designing OTEL

	▲	valyala 6 hours ago \| parent [-]
		These tricks are essential for every database optimized for metrics / logs / traces. For example, you can read on how VictoriaMetrics can compress production metrics to less than a byte per sample (every sample includes metric name, key=value labels, numeric metric value and metric timestamp with millisecond precision). https://faun.pub/victoriametrics-achieving-better-compressio...

▲

pdimitar a day ago | parent | prev [-]

Very curious to read your code doing it. Thought of a very similar approach but never had the time. Are you keeping it somewhere?

▲

jiggawatts 17 hours ago | parent [-]

I only ever got it to a proof of concept. The back end worked as advertised, the issue was that there are too many bugs in WMI so collecting that many performance counters had weird side effects.

Google was doing something comparable internally and this spawned some fun blog titles like “I have 64 cores but I can’t even move my mouse cursor.”

	▲	pdimitar 9 hours ago \| parent [-]
		Ah, I don't mean the Windows-specific stuff. I mean the binary diffing and RLE. While not difficult, I am just curious how others approached it.