What teams use Telflo for

Everyone in the channel agrees the DEBUG logs are junk nobody reads. The filter PR still sits open a month later, because nobody wants to be the one who approved dropping data an incident later needed.

An engineer on Hacker News described what happened after cost pressure switched sampling on at a previous job: at the next incident, "we had no idea if it was even logged."

A one-line edit to a sampling policy lands in a 700-line collector.yaml. CI passes, because the check only proves the YAML parses. The reviewer scrolls, tries to work out which of three pipelines the processor is wired into, and approves. Nobody finds the problem until production, when someone asks where the checkout traces went.

The stock tooling can't do better. Running a pipeline against data and checking the output is an open feature request in the Collector repo, filed in March 2024.

Six weeks into a migration, every pipeline still has two exporters: the old vendor and the new one. Finance wants to know why the old bill hasn't moved. Nobody deletes the old exporter, because nobody can prove the new backend sees everything the old one did. So the migration stalls, and you keep paying for both.

An engineer writes a card-number regex for the redaction processor. The config loads, the collector starts, and nothing anywhere says whether a real card number would come out masked. Someone asked exactly that in the OpenTelemetry repo, full config attached. The discussion has zero replies.

Most teams find out after the fact. One engineer on Hacker News described PII flowing into logs for a month before anyone noticed, then spending three days deleting it record by record.

Your vendor raises the price at renewal, and someone asks what it would take to switch. If the vendor's SDK is embedded in every service, the answer is a re-instrumentation project, so you sign the renewal. Teams adopt OpenTelemetry to escape exactly this: freedom to switch vendors is the second most cited reason for adopting it in Grafana's latest survey.

Some pipeline tools bring the lock-in back by making you write configs in their own proprietary format, one that compiles to OTel instead of being OTel.

Hand-editing collector config means one typo can take the whole pipeline down. You change one setting, restart, and the collector refuses to start, with an error that names no file and no line number. That error has been a known problem since December 2024, so you find the typo by trial and error, which can take hours.

And the files only grow. Real configs run 600 lines and up, and every environment keeps its own slightly different copy. Code reviews of files that size, as one observability shop put it, "become exercises in scrolling." That adds up to a lot of senior engineering time going into a text file.