Analytics SLA Credits: What BigQuery, Redshift, Synapse and Athena Pay When They Fail
September 21, 2026
Data warehouses carry the same promise shapes as everything else in the cloud, plus one twist: three of the four big-platform warehouses measure errors rather than minutes. BigQuery commits to 99.99% for the paid editions and 99.9% for Standard. Redshift splits three ways, from 99.99% multi-AZ down to 99.5% for a single-node cluster. Azure Synapse holds its dedicated SQL pools to 99.9%. Athena commits to 99.9% per region. That spread matters at claim time. On a 30-minute incident, only the 99.99% promises breach, so half the table pays nothing. The ceilings differ just as sharply, from a quarter of the bill on Synapse to the full bill on Redshift and Athena.
The commitments, by platform and deployment shape
| Warehouse | Commitment | What actually counts | Deployment caveat |
|---|---|---|---|
| BigQuery, paid editions | 99.99% | an error rate above 10% in a minute, HTTP 500s among valid requests, minimum 20 requests | Standard edition drops to 99.9%; Omni is excluded entirely |
| Redshift, Multi-AZ | 99.99% | minutes in which all connections to the cluster fail | needs two or more compute nodes in each of two or more availability zones, on the current or trailing maintenance track |
| Redshift, single-AZ multi-node or Serverless | 99.9% | same | one availability zone, two or more nodes |
| Redshift, single-AZ single-node | 99.5% | same | one compute node |
| Azure Synapse, dedicated SQL | 99.9% | minutes in which more than 1% of client operations return an error code | measured per database |
| Amazon Athena | 99.9% | the average of 5-minute intervals in which requests fail with a 500 or 503 | only StartQueryExecution calls are in scope; empty intervals count as 100% available |
BigQuery carries one more quirk: its Data Transfer Service runs on a delivery promise instead of an uptime one, committing to scheduled loads within 24 hours and refunding the day's transfer charges when it misses. The delivery SLO applies only to automatically scheduled runs.
Error rates, not stopwatches
The measurement rules are where these SLAs diverge from compute and network promises, and they decide claims more often than the headline numbers.
- BigQuery counts server-side errors: valid requests returning HTTP 500s, above ten percent of traffic within a single minute, with a floor of 20 valid requests. Repeated identical requests do not count toward the rate at all unless the client backs off, one second doubling to 32.
- Redshift is the only one of the four measuring connectivity in the traditional sense: a minute is unavailable when every connection to a running cluster fails during it.
- Synapse counts client operations, and a minute tips into downtime once more than one percent of them return an error code.
- Athena averages per-request failures over 5-minute windows. An interval in which nobody queried counts as fully available, and only the StartQueryExecution API is covered.
Two consequences follow. A warehouse that is up but pathologically slow never breaches BigQuery, Synapse or Athena, because their clocks only tick on errors. And on all four, silence reads as health: no traffic in a measurement window is treated as perfect availability.
Three of these four promises pay when requests fail, not when time is lost. A slow query, a hung console, a dashboard that takes forty seconds to render: none of it counts. The claimable event is an error rate, and the evidence is a log of failures, not a stopwatch.
The credit ladders
| Warehouse | First tier | Middle tier | Top tier |
|---|---|---|---|
| BigQuery, all editions | 10% (99.0% to below the commitment) | 25% (95% to below 99%) | 50% (below 95%) |
| Redshift, all three shapes | 10% | 25% | 100% |
| Synapse SQL | 10% (below 99.9%) | 25% (below 99%) | no top tier |
| Athena | 10% | 25% | 100% |
Synapse is the stingiest of the four and it is not particularly close: dedicated SQL pools have no 100% tier, so even a catastrophic month pays 25%. Fabric, the successor platform for new analytics work, carries the same two-tier shape. Google caps every payout at half the bill, and applies approved credits within 60 days of the request. AWS pays up to the entire monthly warehouse charge, and the single-node Redshift shape holds one of the loosest commitments on the list: 99.5% tolerates more than three and a half hours of failed connections in a month before the ladder even starts.
Money math on a $1,000 monthly warehouse bill
A 30-day month holds 43,200 minutes, and 30 minutes of outage puts the month at 99.931%.
| Incident | BigQuery | Redshift Multi-AZ | Synapse SQL | Athena |
|---|---|---|---|---|
| 30 minutes (99.931%) | $100 | $100 | $0 | $0 |
| 8 hours (98.889%) | $250 | $250 | $250 | $250 |
| 40 hours (94.444%) | $500 | $1,000 | $250 | $1,000 |
Read the first row, where the deployment shape does the talking: only the 99.99% promises breach, so the paid editions of BigQuery and the multi-AZ Redshift shape pay while the 99.9% commitments keep their money. A single-node Redshift cluster collects nothing at 30 minutes either, and BigQuery Standard edition tracks Synapse and Athena. Read the last row, where the ceilings do the talking: AWS pays the whole bill, Google's cap holds it to half, and Synapse never gets past a quarter. The same outage, the same bill, answers that differ by a factor of four.
Filing inside the windows
| Warehouse | What the claim needs | Deadline |
|---|---|---|
| Redshift | "Redshift SLA Credit Request" in the subject; dates, times, cluster names and regions; request logs; the three shapes cannot be stacked on one deployment | end of the second billing cycle |
| Athena | "SLA Credit Request"; dates, times and per-5-minute-interval availability for the region, not just a status page screenshot | end of the second billing cycle |
| BigQuery | a technical support contact with log files showing the error periods; claims are per project | 30 days from eligibility |
| Synapse | a portal support request with the incident description, its duration and the affected resources | within two months of the billing month's end |
AWS pays credits only once they pass one dollar, and returns confirmed claims within a billing cycle. Google's clock is the shortest of the set, 30 days, and the claim evidence guide covers what each reviewer accepts.
The audit worth running: list every warehouse, its deployment shape, and the commitment that shape earns, then write the measurement rule beside it, because that is the rule applied to your logs. An Athena claim needs interval-level availability data. A BigQuery claim needs evidence of 500 rates across a minute. A Redshift claim needs failed connections. The right outage with the wrong record still pays nothing, and this is the interval-level evidence UptimeAudit assembles when it watches the big four's health feeds and drafts claims against what you actually run.