Should an internal platform have SLOs, and what should they cover?

Yes, and the reason is specific

Consumers build on the platform's availability, so the platform's reliability becomes theirs. A team whose service depends on a platform component inherits its failure modes, and without a stated SLO they cannot reason about their own — they are building on an unknown.

An unstated platform SLO means every consumer assumes a different one, and the assumptions are usually optimistic.

What the SLOs should cover

• The data path separately from the control plane. Customer traffic served through platform components is a different reliability question from the ability to deploy — and the data path's target should be far higher.
• Deployment success rate and duration, since a platform that deploys unreliably is a platform teams work around.
• Build and pipeline availability, which blocks all delivery when it fails.
• Provisioning latency, which is the platform's user-facing promise for self-service.
• Observability pipeline availability, because losing telemetry during an incident is a compound failure.

The crucial architectural implication

A platform whose control plane failure stops customer traffic has capped every consumer's availability at its own. That is rarely what anyone intended and it is the default unless static stability is designed in: configuration cached locally, no synchronous platform call on the request path, and a failure mode of "serve the previous configuration".

The platform SLO for the data path should therefore be achievable through independence rather than through the platform being extremely reliable.

The peak-event consequence

For a platform serving a workload with extreme scheduled peaks, the platform's own control plane is under its maximum stress at exactly the same moment, because scaling activity and configuration changes cluster around the event. Depending on it during the peak correlates the two failures.

What the SLO changes practically

It makes the platform's reliability a negotiable, budgeted thing rather than an assumption — and it gives consumers a number to design against, which is what allows them to build their own degradation correctly.