Why I Built Sampleless
After a decade building and operating cloud infrastructure, I reached a breaking point. The tools we use to understand our systems were actively working against us.
The moment it clicked
It was 2AM during an incident. We were trying to trace a payment failure affecting a small but important segment of customers. The traces we needed were not there.
Not because we had not instrumented the service. Not because the traces had expired. They simply had not been sampled.
At 1% sampling, with these customers representing maybe 0.5% of traffic, the probability of capturing their transactions was effectively zero. We had paid for observability and gotten blindness.
That night, bleary-eyed and frustrated, I did the math. At our volume, the cost difference between 1% and 100% sampling was not the storage. It was the egress. Shipping 12TB/month to a SaaS provider meant $18,000-96,000 annually in networking costs alone, before the observability subscription.
What if the data never left?
Ten years of watching the problem grow
I have spent over a decade in cloud infrastructure. AWS, Azure, Kubernetes at scale. I have been the person woken at 3AM and the person who had to explain to executives why the outage lasted four hours instead of forty minutes.
Every company I worked at had the same pattern:
- Start with full observability during development
- Hit production scale and watch costs explode
- Implement sampling to control costs
- Lose visibility into rare but important events
- Spend incidents hunting for data that was never collected
The industry accepted this as normal. I stopped accepting it.
The economics are backwards
Traditional observability pricing creates perverse incentives at every level:
Per-host pricing means your architecture decisions are influenced by your monitoring bill. Do you really want fewer, larger instances because of observability costs?
Per-GB pricing means you log less. You sample more. You actively discard the data you might need tomorrow.
Custom metrics surcharges mean you avoid instrumenting business metrics. The most valuable signals become the most expensive to collect.
And underneath all of this, cloud providers charge egress fees that make high-volume telemetry economically impossible to ship off-premises.
The result: companies pay premium prices for partial visibility, then pay again in incident duration when the data they need is not there.
What if data never left your cloud?
The insight that led to Sampleless was simple: if telemetry data stays in your cloud account, egress costs disappear.
At $0.09-0.12/GB for internet egress, a company generating 10TB/day of telemetry pays over $450,000 annually just in networking costs. That is before the SaaS subscription, before storage, before anything else.
BYOC—Bring Your Own Cloud—changes the equation entirely. The observability platform runs in your VPC, processes your data in-place, and stores it in your storage. The egress cost drops to zero.
Suddenly, 100% collection becomes economically viable. Not as a premium tier. As the default.
Why ML needs complete data
There is a deeper reason to care about full-fidelity collection beyond debugging individual incidents.
Machine learning models are only as good as their training data. Anomaly detection trained on sampled data learns from an incomplete picture. It cannot detect anomalies in patterns it has never seen.
Datadog Watchdog requires two weeks of historical data to train metric baselines. If that data is 1% sampled, the baselines are biased toward high-volume patterns. Rare but important behaviors are invisible.
We built ALBA—Adaptive Learning Behavioral Analytics—to take advantage of complete data. Entity-level behavioral scoring that aggregates anomaly detection with business impact weighting. It works because we have the full picture.
Sampling is not just a cost problem. It is a capability problem.
What I am not claiming
I want to be direct about what Sampleless is and is not.
Datadog has over 1,000 integrations. They have built an incredible platform over a decade with thousands of engineers. We do not have 1,000 integrations. If you need extensive pre-built connectors to legacy systems, Datadog might be a better fit.
We are not trying to be a "better Datadog." We are taking a fundamentally different approach. OpenTelemetry-native instead of proprietary agents. BYOC instead of SaaS. Flat pricing instead of per-host and per-GB.
That approach has tradeoffs. We believe those tradeoffs are worth it for teams that prioritize complete visibility over integration breadth.
No vendor lock-in
One decision we made early: everything we build would use open standards.
Sampleless is OpenTelemetry-native. Your instrumentation is not proprietary. If you decide to leave, your traces, metrics, and logs export to any OTLP-compatible backend.
ALBA, our behavioral analytics engine, is built on OpenALBA—an open specification for entity-level scoring. We believe the industry needs shared standards for behavioral analytics, not another proprietary lock-in vector.
This is not just philosophy. It is practical. Teams should evaluate observability platforms on capability, not switching costs.
The road ahead
Building Sampleless has been the most challenging and rewarding work of my career. Every week, I talk to engineering leaders frustrated by the same problems I faced. The observability tax. The sampling compromise. The incident where the trace was not there.
We are not done. There is an enormous amount of work to do. But the foundation is solid: an architecture that makes full-fidelity collection economical, pricing that does not penalize you for using the product, and open standards that respect your autonomy.
If you have ever stared at a sampling configuration trying to balance cost and visibility, I built Sampleless for you. Let me know what you think.
See the approach in action
Book a demo to see how BYOC architecture changes the economics of observability.