Scraping API SLA: What to Get in Writing Before You Sign
An uptime number on a pricing page is not an SLA. Here are the seven terms a data API contract should spell out when your product depends on it: uptime scope, failed-request billing, sustained throughput, price ceilings and more.
Most data API vendors print an uptime figure somewhere on their site. Far fewer put one in the document you sign, and fewer still say what the number covers. If a dashboard, a model or a customer-facing feature depends on the data arriving, the gap between “99.9% uptime” in a hero banner and an actual commitment is where your on-call nights come from.
This is the checklist we would want if we were buying: seven terms to get in writing, what a good answer looks like, and the questions that expose a weak one.
1. Uptime: of what, measured how, over what period
“99.9%” on its own tells you nothing until three things are attached to it:
- Scope. Uptime of the vendor’s network, of the API gateway, or of the specific platform you buy data from? A gateway that answers
200with an empty body while one platform is broken is “up” by some definitions. - Period. Monthly is the standard. A yearly figure lets a vendor spend a whole bad week and still hit the number.
- Who measures. Ask whether you get a status page that shows your platforms, not a global page where your outage disappears into an average.
What the arithmetic means in practice:
| Monthly uptime | Downtime allowed in a 30-day month |
|---|---|
| 99.0% | about 7 hours 12 minutes |
| 99.5% | about 3 hours 36 minutes |
| 99.8% | about 1 hour 26 minutes |
| 99.9% | about 43 minutes |
| 99.95% | about 22 minutes |
The difference between 99.0% and 99.95% is roughly seven hours a month of your pipeline waiting. Ask for the number per platform you depend on, monthly, with a status page you can point your own team at.
2. Failed requests: who pays for them
This is the clause that matters most on a normal day, because data APIs fail partially far more often than they go fully down. A platform changes a page, a profile goes private, a request times out. The question is simple: when a request fails, are you billed for it?
A good answer covers three cases explicitly:
- server errors and timeouts on the vendor’s side: never billed;
- requests that return nothing usable because the vendor couldn’t get the data: never billed;
- your own bad input (a malformed URL, a missing parameter): ideally not billed either, since it costs the vendor almost nothing to reject.
If the contract is silent, assume failures bill. At high volume, a 5% failure rate on a metered plan is 5% of your bill paid for nothing, and it gets worse exactly when the platform is having a bad week.
3. Throughput: sustained, not “up to”
Rate limits on pricing pages are usually ceilings: “up to 50 requests per second”. A ceiling is not a guarantee. Under load from other customers you may get a fraction of it, and that is precisely when you need it.
What to ask for instead:
- a sustained rate, in requests per second, that is reserved for you on the platforms you depend on;
- what happens when you go above it: queued, rejected with
429, or absorbed when there is room; - whether that capacity is shared with other customers or set up for your account.
A monthly request quota and a guaranteed rate are different promises. A quota tells you how much you may use; a rate tells you how fast you will actually get it. If you run a daily batch in a two-hour window, only the rate matters.
4. Price: held, with a ceiling for after
Data API pricing moves. Platforms change, vendors reprice, introductory plans get retired. For a dependency you plan around, ask for:
- a price guarantee period (12 months is reasonable) for the workload you describe up front;
- a ceiling in writing for after it ends: the most the same volume can ever cost you;
- whether growing (more volume, another platform) reopens the price of what you already have. It shouldn’t.
A price guarantee that also locks you into a 12-month term is a trade, not a gift. The better version binds the vendor and leaves you month to month.
5. Support: a channel and a response time
“Priority support” is not a term. Ask which channel (email queue, shared Slack or Telegram channel, phone), what hours, and who answers: a ticket triage team or the engineers who run your setup. For a pipeline that runs overnight, round-the-clock coverage in a channel your engineers already watch is worth more than a fast response SLA on a ticketing portal nobody checks at 3am.
6. Change management: what happens when a platform changes
Social and e-commerce platforms change their pages and apps constantly. That is the real reliability risk, more than server uptime. Ask:
- how quickly a broken endpoint is fixed, and whether you are told;
- whether field names and response shapes are stable, and how changes are announced;
- whether you can request a new field or endpoint, and how long that takes for a customer at your volume.
7. Data handling: what the vendor keeps
Your security review will ask this, so get it before they do: are your queries (the accounts, keywords and URLs you look up) logged, and for how long? Is returned data cached or resold? Which legal entity are you contracting with, and under which data protection law?
The questions, in one place
Copy these into your vendor email:
- What monthly uptime do you commit to, in writing, for the platforms we use? Is there a status page per platform?
- Are failed requests (errors, timeouts, no data returned) billed?
- What sustained requests per second is reserved for us, and is it shared with other customers?
- How long is our price held, and what is the ceiling after that?
- What support channel and hours do we get, and who answers?
- How are platform changes handled and communicated?
- Do you keep our query parameters or the data you return?
How we answer them
We publish our answers on our enterprise page, but here is the short version for teams on a dedicated plan:
- 99.95% monthly uptime on your dedicated platforms (about 22 minutes a month), with your own private status page. If a component fails, your keys keep working.
- Failed requests are never billed: errors, timeouts and bad-input 400s cost nothing.
- Guaranteed sustained throughput, sold in blocks of about 21 requests a second (roughly 55 million requests a month if run non-stop) on the platform they are bought for. No monthly cap.
- Price held for up to 12 months with a ceiling in writing for after, while the plan stays month to month.
- 24-hour support in a private Slack or Telegram chat with the people who run your setup.
- New fields and endpoints on request, built first for what is needed at volume.
- Your queries are not kept when a request succeeds, and returned data is not cached or stored. Details are in our privacy policy.
If you would rather walk through your workload than read a page, book 15 minutes and we will tell you what it needs. Sometimes the honest answer is that pay-as-you-go is enough.
Frequently asked questions
What is a good uptime SLA for a scraping or data API?
For a production dependency, 99.5% to 99.9% monthly, scoped to the platforms you actually use. More important than the decimal is that it is written into your terms, measured monthly and visible on a status page.
Is a success rate the same as uptime?
No. Uptime says the API answered. Success rate says it returned the data you asked for. A vendor can be “up” while a platform returns errors. That is why non-billing of failed requests matters: it puts the cost of a low success rate on the vendor instead of on you.
Should a data API SLA include a guaranteed request rate?
If your workload is time-bound (a nightly batch, near-real-time monitoring), yes. A monthly quota does not tell you how fast you can pull it. Ask for a sustained requests-per-second figure reserved for your account.
Do I need a long-term contract to get these terms?
No. A price guarantee can bind the vendor while you stay month to month. Be wary of terms where the only way to get a price lock is a 12-month commitment from you.