Skip to content
SRE retainer guide · checked 13 November 2026

SRE as a service, with the contract terms written out

What a site reliability retainer should buy, how to size it, which response targets to agree, how SLOs and error budgets belong in the contract, how follow-the-sun on-call from India works, and how you leave with everything you paid for.

IT engineering team collaborating on laptops
At a glance
  • SLOs in the contractIndicators, windows and error-budget policy written down
  • Follow-the-sun from IndiaYour night hours covered by engineers in their working day
  • Hours for fixingEngineering time reserved, not only incident response
  • Exit built inRunbooks, alert rules and dashboards live in your repositories

What SRE as a service means

SRE as a service is a retainer under which an external team takes shared responsibility for a system’s reliability: it agrees service level objectives with you, watches error budgets, carries the on-call pager for an agreed window, runs incidents, writes postmortems and spends reserved engineering time removing the causes and the toil. It differs from managed services, which keep systems running to a ticket SLA, because the provider is measured on your users’ experience and is expected to change the system, not just operate it.

Managed SRE services from India

Most “SRE as a service” offers are on-call with a new name. Here is how to tell the difference

Page one for this query is mostly vendor pages that list the same words (SLOs, observability, incident response) without saying what the monthly fee buys. The useful questions are commercial. How many engineering hours are in the fee, and how many of them go to fixing things rather than answering pages? Which hours of the day does the pager cover, and how fast must someone respond? Whose number is the SLO, and what happens when the error budget runs out? What do you keep if you end the contract?

This page answers those questions for the way we sell it, and gives you a checklist to use with any provider. The practice itself, defining indicators, instrumenting with OpenTelemetry and redesigning alerts, is described on our SRE and observability page. If what you need is someone to patch servers and restore backups under a ticket SLA, our managed services are the better fit, and for AWS-specific operations in India see our AWS managed services guide.

Radiatus is an engineering studio registered as an MSME in India and in Delaware, USA. Our engineers work from India, which matters for one reason above all: for teams in the Americas and much of Europe, your night falls in our working day.

SRE as a service, managed services, staff augmentation or your own team?

Four models get sold under overlapping names. They differ in who owns the reliability target and who is allowed to change the system.

ModelMeasured onCarries the pagerChanges code and infrastructureTypical failure
SRE as a serviceSLOs on user journeys, error-budget burn, postmortem actions closedProvider, for an agreed windowYes, through your review processEngineering hours quietly consumed by interrupts
Managed servicesTicket response and resolution times, patch and backup complianceProviderInfrastructure only, rarely application codeSystems kept running, causes never removed
Staff augmentationWhatever your manager setsYour rota, which they joinYes, as your team memberOne person becomes a single point of failure
In-house SRE teamYour own objectivesYour teamYesToo few people for a sustainable 24/7 rota

A sustainable round-the-clock rota needs enough people that nobody is on call too often, which is why small companies rarely staff it themselves. A retainer shares one rota across clients, but that is also its risk: ask how many systems the person holding your pager also covers.

Example tiers to agree, sized in engineer-days

We have not published fixed tiers or response numbers, because the right figures depend on your architecture, how much you let us change and how noisy your alerts are on day one. The table shows the shapes we propose and the response targets we would expect to agree. Treat every figure here as an example for discussion; the binding numbers are the ones written into the contract after onboarding.

Business hoursExtendedRound the clock
Engineering capacityAbout 4–6 engineer-days a monthAbout 8–12 engineer-days a monthAbout 12–20 engineer-days a month, plus the on-call rota
On-call windowIndian working day, Monday to FridayTwo Indian shifts, about 16 hours on weekdays; weekend pages next morning24 hours, 7 days
Example Sev 1 target (users cannot use a core journey)Acknowledge within 30 minutes in windowAcknowledge within 15–30 minutes in windowAcknowledge within 15 minutes, any time
Example Sev 2 target (degraded, workaround exists)Within 2 working hoursWithin 1 hour in windowWithin 1 hour, any time
Example Sev 3 (no user impact yet)Next working dayNext working dayNext working day
SLOs under watch2–3 user journeysUp to about 6Agreed per service
ReviewsMonthly reliability reportMonthly report and review callMonthly review, quarterly SLO review with product owners
SuitsTeams in Asia, the Gulf or Europe with few night incidentsTeams in Europe or the Americas wanting their night coveredRevenue or clinical systems that must be watched every hour

“Acknowledge” means a named engineer has the incident and has posted a first assessment, not that a bot replied. Ask any provider for the same distinction. Resolution times are not promised in any tier, because they depend on causes outside the provider’s control; what is promised is the response, the communication rhythm and the follow-up.

What other providers publish: one price reference

Few SRE-as-a-service providers publish prices. Palark’s pricing page does, and it is a useful benchmark for the shape of an offer. As of 13 November 2026 it lists SRE as a Service at €9,100 a month for 120 team hours (M), €11,700 for 160 hours (L) and €14,300 for 200 hours (XL), each with 24/7 on-call, unlimited incidents and a 10-minute reaction-time SLA, and an application uptime SLA that is “TBD after 3 months”. Its DevOps as a Service tiers, €3,200 for 40 hours and €6,200 for 80 hours, cover 8×5 rather than 24/7. That works out at roughly €71–76 per team hour at the SRE tiers. Two things in that offer are worth copying whatever you buy: hours are stated in the price, and the uptime commitment waits until the provider has seen the system for three months.

We quote in engineer-days per month after onboarding, not from a public price list, and we tell you how many of those days we expect to go to engineering rather than response.

How SLOs and error budgets go into the contract

A provider cannot promise your uptime for a system it does not control. What a contract can do is define how reliability is measured and what both sides do when it slips. Write down five things:

  1. The indicator. The exact query, the data source and what counts as good. “Share of booking API requests that return non-5xx in under 800 ms, measured at the load balancer” is checkable; “the site is up” is not.
  2. The objective and window. For example 99.9% over a rolling 28 or 30 days.
  3. The error-budget policy. Google’s example policy halts changes other than P0 and security fixes while a service is over budget, and requires a postmortem with at least one P0 action for any single incident that uses more than 20% of the four-week budget. Your policy can be softer, but it must say who decides.
  4. Exclusions. Planned maintenance, a cloud provider’s regional outage, load tests, traffic from clients outside the agreement.
  5. Who changes the objective. Usually a quarterly review with product owners, never mid-incident.

The arithmetic of a 30-day window, so everyone sees what a target allows:

ObjectiveError budget over 30 daysWhat that usually implies
99%7 hours 12 minutesBusiness-hours cover can be enough
99.5%3 hours 36 minutesNight pages for core journeys
99.9%43 minutes 12 secondsRound-the-clock on-call and fast rollback
99.95%21 minutes 36 secondsAutomated failover; humans too slow for most faults
99.99%4 minutes 19 secondsRedundancy across zones or regions, designed in

Keep the objective as code next to your services. An example in the format used by Sloth, which generates Prometheus recording and alerting rules from it:

version: "prometheus/v1"
service: "booking-api"
labels:
  owner: "sre-retainer"
slos:
  - name: "requests-availability"
    objective: 99.9
    description: "Booking API requests answered without a 5xx."
    sli:
      events:
        error_query: sum(rate(http_requests_total{job="booking-api",code=~"5.."}[{{.window}}]))
        total_query: sum(rate(http_requests_total{job="booking-api"}[{{.window}}]))
    alerting:
      name: BookingApiErrorBudgetBurn
      page_alert:
        labels:
          severity: page
      ticket_alert:
        labels:
          severity: ticket

Generated alerts follow the multi-window, multi-burn-rate method from Google’s SRE workbook: page when the budget is burning 14.4 times too fast over an hour (and over the last five minutes), or 6 times over six hours; open a ticket at a steady 1 times over three days. That one change usually removes most CPU and disk pages that never meant a user was affected.

On-call: write down what the provider may do without asking

Most on-call friction between a company and an outside SRE team comes from one gap: nobody agreed which actions are pre-approved. Settle it in a table before the first shift.

Pre-approved during an incidentNeeds your on-call contact’s approvalNever without a change request
Restart a process or pod; scale out within agreed limits; roll back the most recent deploy; fail over to a standby; switch on a documented feature flagScale beyond the cost limit; fail over across regions; block a customer or IP range; disable a scheduled jobData fixes or deletions; database schema changes; changes to security groups, IAM or encryption keys

Pair that with access rules: personal accounts with multi-factor sign-in, elevated access granted for the incident and expiring after it, and session logging your security team can read. Shared admin passwords end the conversation.

Check your paging tool too. Atlassian stopped selling Opsgenie on 4 June 2025 and ends support on 5 April 2027, and Grafana OnCall’s open-source edition was archived on 24 March 2026. If your rota runs on either, moving it is part of onboarding, and a good moment to rebuild escalation policies from scratch.

Follow-the-sun from India: which of your night hours we cover awake

India runs on IST, UTC+5:30, with no daylight saving. Rather than overlap for meetings, the question for on-call is which hours of your night fall inside an Indian working day:

  • US East Coast: 9 p.m. to 7 a.m. Eastern in winter is 7:30 a.m. to 5:30 p.m. IST. The whole of your night sits in our day shift.
  • US West Coast: 9 p.m. to 7 a.m. Pacific in winter is 10:30 a.m. to 8:30 p.m. IST, again inside a normal day plus an evening shift.
  • UK: 10 p.m. to 7 a.m. GMT is 3:30 a.m. to 12:30 p.m. IST. A morning shift from 7 a.m. IST covers the second half of your night; the first half needs a night shift or your own rota.
  • Gulf and Singapore: your night is our night too. Round-the-clock cover from India for these zones is a genuine night shift, priced accordingly.

The handover between shifts is where incidents get dropped. Each one should be a written note in the incident channel with four lines: incidents open and who owns them, error budgets burning faster than normal, changes in flight or scheduled, and anything the next shift must not do.

Runbooks: one per alert, tested

An alert without a runbook is a question sent to a sleepy engineer. Every paging alert gets one, kept in your repository and linked from the alert itself:

# BookingApiErrorBudgetBurn
Meaning:    booking API 5xx rate is using the monthly error budget too fast
User impact: patients cannot complete online bookings
First checks:
  1. Deploys in the last 2 hours?  (pipeline history)
  2. Database connections saturated?  (dashboard: booking-db)
  3. Upstream payment or SMS provider status page
Safe actions: roll back last deploy; scale booking-api to 6 replicas
Escalate:   to your on-call contact if no improvement in 20 minutes
Last tested: 2026-10-14 (game day)

We test runbooks in game days, deliberately breaking a staging copy, and any step that failed or confused someone is fixed before the runbook is marked tested.

Postmortems that change the system

Agree in the contract when a postmortem is due (for example, every Sev 1 and any incident that used more than a fifth of the monthly budget), how soon a draft follows the incident, and who reviews it. The format matters less than three rules: the timeline is built from logs and chat, not memory; the causes describe conditions, not people; and every action item has an owner, a date and a ticket. The number that tells you whether a retainer is working is not the count of postmortems written. It is the share of their action items closed by the due date, reported every month.

Toil: count it, cap it, spend the rest fixing

Google’s SRE book defines toil as work that is manual, repetitive, automatable, tactical, has no enduring value and grows with the service, and states a goal of keeping it below 50% of each SRE’s time. The same principle belongs in a retainer. We tag every hour as either interrupt (pages, tickets, requests) or engineering (automation, alert fixes, capacity work, postmortem actions) and show the split in the monthly report. If interrupts take more than half of the retained days for two months running, that is a finding, and the review decides whether to fix the noisiest sources first or change the tier.

Exit and handover

A provider you cannot leave is a risk, whatever its response times. Write exit into the first contract:

  • Everything lives in your accounts. Alert rules, SLO definitions, dashboards and runbooks as code in your repositories; postmortems in your wiki; tickets in your tracker.
  • A notice period and a handover window. For example 30 days’ notice and two weeks of shadow on-call, with your team or the new provider holding the pager and ours as second line.
  • An access list. Every account, key and integration the provider holds, maintained monthly, so revocation on the last day is a checklist rather than an investigation.
  • Credential rotation. Any shared secret the provider could see is rotated after exit.

Questions to ask any SRE as a service provider

  1. How many engineering hours are in the fee, and what share went to engineering rather than interrupts for your current clients last quarter?
  2. How many systems does the engineer holding our pager also cover at night?
  3. Do your response targets mean a human acknowledgment with a first assessment?
  4. Whose SLOs are these, how are they measured, and what happens when the budget is spent?
  5. What may your on-call engineer do without asking us?
  6. When is a postmortem due, and how do you report the action items closed?
  7. What do we keep, and how long is the handover, if we leave?
Packages

SRE as a service: how we package it

Sized in engineer-days. Figures are typical; we fix the retainer after the onboarding baseline, when we know how noisy the system is.

Reliability baseline (onboarding)

The problem

You want an outside team on call, but nobody can say what “reliable” means for your system or how noisy its alerts are.

How we do it

We review recent incidents, alert volume and existing dashboards, agree two to six user journeys, write their indicators and objectives, rebuild paging alerts as burn-rate alerts, and write a runbook for each. Typical effort: 6–12 engineer-days, once.

What you get

  • SLO specification as code
  • Alert inventory with keep, fix or delete for each
  • Runbook per paging alert
  • Proposed tier and on-call permissions table
Onboarding

How an SRE retainer starts

From signature to a full rota in about a month. Nobody takes your pager on day one.

Week 1

Access and listening

Read-only access, incident history and alert export; interviews with whoever holds the pager today.

Weeks 1–2

Baseline

User journeys and SLOs agreed, alerts rebuilt as burn-rate alerts, runbooks written, permissions table signed.

Week 3

Shadow on-call

We shadow your rota: we see every page and act only when asked.

Week 4

Reverse shadow

We hold the pager with your engineer as backup; handovers practised.

Month 2 on

Retainer

Full on-call for the agreed window, monthly report and review, engineering backlog worked.

Senior engineers firstYour first call is with someone who will do the work, not a sales script.
You own everythingCode, configuration, reports and documentation stay in your accounts and are handed over.
We run our own products15 live platforms we build and operate, so we feel production pain too.
Honest about fitIf a cheaper route or another provider suits you better, we will say so.
FAQ

Questions about SRE as a service

A retainer under which an external team shares responsibility for a system’s reliability: it agrees service level objectives, watches error budgets, holds the pager for an agreed window, runs incidents and postmortems, and spends reserved engineering time removing causes and toil. Managed services, by contrast, usually keep systems running to a ticket SLA without changing them.

It depends on engineering hours and the on-call window. For reference, Palark publishes SRE as a Service at 9,100 to 14,300 euros a month for 120 to 200 team hours with 24/7 on-call, as of 13 November 2026. We size retainers in engineer-days a month after an onboarding baseline, typically 4 to 12 days for business-hours or extended cover and 12 to 20 plus the rota for round-the-clock.

Response, not resolution. A reasonable starting point to negotiate is a human acknowledgment with a first assessment within 15 minutes for a severity-one incident in a 24/7 tier, and within an hour for degraded service. The binding numbers belong in the contract and should be reviewed after the first quarter, once alert noise is known.

Not honestly at the start, because it does not control your code, your releases or your cloud provider. What it can commit to is how reliability is measured, how fast it responds and what happens when the error budget is spent. Some providers set an uptime commitment only after observing the system for a few months.

India is on UTC+5:30 with no daylight saving, so a US East Coast night falls almost exactly in an Indian day shift and a West Coast night in a day plus evening shift. Engineers handle your night pages awake. For the UK, a morning shift covers the second half of your night; Gulf and Singapore nights are Indian nights too and need a true night rota.

You do. They should live as code in your repositories, your wiki and your tracker from the first day, so that ending the contract means revoking access, not recovering documents. Ask any provider to put this in writing.

Whatever your written error-budget policy says. A common pattern, from Google’s published example, is to pause feature releases except urgent and security fixes until the service is back within its objective, and to require a postmortem for any single incident that used more than a fifth of the budget.

Give the agreed notice, then run a handover window, typically two weeks of shadow on-call where your team or the new provider holds the pager and the outgoing team is second line. Close by revoking every account on the maintained access list and rotating any shared secrets.

Last updated