New teams either monitor nothing or monitor everything. Both fail. Nothing means customers are your detection layer. Everything means the pager fires for pages nobody will fix tonight, and soon nobody believes the pager at all.
The day-one list
Put external checks on:
- Login or session bootstrap — if people cannot get in, nothing else matters.
- The main app shell or primary API health URL — whatever "the product works" means in one request.
- Checkout, billing portal, or payment webhook receiver — if money touches it, watch it.
- Your public status page — on separate infrastructure if you can. When the product is down, this is how you talk.
That is enough for an early SaaS. Four good monitors beat forty mediocre ones.
What to skip until it hurts
- Marketing homepage only, if login and API are already covered and the marketing site is static elsewhere.
- Every internal admin route.
- Every microservice behind the gateway you already check.
- Staging, unless someone gets paid to wake up for staging.
- Third-party status URLs you do not own — subscribe to their status page; do not invent a second pager for Stripe's incidents unless you have a real playbook.
If a check fails and the runbook is "shrug, wait for the vendor," it should not wake a human.
How to grow the list
Add a monitor when you can finish this sentence: If this breaks at 2am, the on-call person will ___.
Ship a new critical customer flow → add it.
Get burned by an SSL expiry → monitor certs or the HTTPS URL that fails when they expire.
Ship a cron that must run nightly → add a heartbeat once you have one.
Enterprise customer asks for component-level status → split status page components and back each with a check.
Do not add monitors because a competitor's marketing page shows a big number.
One rule that keeps the list honest
Every monitor should have an owner and a next step. Owner can be "whoever is on-call." Next step can be a three-line runbook. No owner and no next step means the alert is decoration.
On Tallwatch I would rather you use ten free monitors on the paths that matter than fill the board with noise. Coverage is not the same thing as care.