System Design Interviews in the UK: Resilience and Numbers

A timed method for UK system design rounds, with a payments design sized against an impact tolerance, the UK GDPR 72-hour clock and the single AWS London region.

By · · 7 min read

Illustration for the System Design Interview Framework guide

United Kingdom edition. Written for UK interviews, especially at banks, insurers and payment firms: FCA and PRA operational resilience (important business services and impact tolerances), UK GDPR Article 33, and UK-only hosting when a provider, such as AWS, has one London region. US edition

A UK system design round runs on the same four steps as anywhere: agree requirements and numbers, sketch the main path, go deep on one or two parts, then talk about scale and trade-offs. What differs is the follow-up questions. At a bank, building society, insurer or payments firm, "how does this fail?" is not a closing pleasantry; it is the centre of the interview, because those firms work under the FCA and PRA operational resilience rules. This edition shows how to fold those rules into your requirements, then works a payments design through with numbers you can put on the whiteboard. The practice prompts and topic list are in the System Design Interview Framework (UK).

The method, timed for a 45 to 60 minute round

StepMinutesOutput on the board
Requirements and estimates8 to 10Core features, non-goals, throughput, storage, availability stated as time
High-level design10 to 15Main request path, APIs, data model
Detailed design15 to 20One or two components: schema, partitioning, failure handling, idempotency
Resilience and scale8 to 10What breaks at peak, how the service degrades, how you recover and reconcile

Say the plan out loud in the first minute so the interviewer can steer you. In UK financial services interviews the last row can take longer than you plan, so leave room for it.

The UK requirement to raise early: impact tolerance

The FCA's operational resilience page explains that its rules came into force on 31 March 2022 and that firms had until 31 March 2025 to make sure they could operate their important business services within their impact tolerances. The rules cover banks, building societies, PRA-designated investment firms, insurers, recognised investment exchanges, enhanced scope SM&CR firms, payment and e-money institutions, consolidated tape providers and qualifying cryptoasset firms.

The PRA's supervisory statement SS1/21, chapter 3, defines an impact tolerance as the maximum tolerable level of disruption to an important business service, measured by a length of time in addition to any other relevant metrics. Three details matter in a design:

  • It always includes time. "Back within 2 hours" is the shape of the requirement, not "99.99% uptime".
  • It can add other metrics, such as the volume or value of transactions affected.
  • It applies at peak, as well as in normal circumstances, so size against the peak rate.

The firm chooses the tolerance; the regulator does not set a number. In an interview, ask for one or propose one and say why.

Worked example: a payments API for a UK bank

Prompt. Design the service that accepts outbound customer payments from a bank's app and passes them to the payment scheme.

Requirements. Accept a payment, validate it, debit the account, submit it to the scheme, report status. Non-goals: fraud scoring internals, statements. Non-functional: no payment lost or sent twice; the firm treats "make a payment" as an important business service with a 2-hour impact tolerance (interviewer's premise).

Estimates (premises you would agree with the interviewer):

volume:   40 million payments a day
average:  40,000,000 / 86,400 ≈ 463 per second
peak:     1,500 per second (salary day, assumed)
record:   ~1 KB per ledger entry
storage:  40 GB a day, ~120 GB with three copies

The tolerance turns into a queue size. If the scheme connection is down for the full 2 hours at peak, the backlog is 1,500 x 7,200 = 10.8 million payments, about 10.8 GB of messages. That decides three things on the board:

  1. A durable queue sized for millions of messages, not thousands.
  2. Idempotency keys on every payment, because customers will retry and the app will resubmit; a replay must not send a payment twice.
  3. Drain time. If the submission tier can push 3,000 payments a second to the scheme once it is back, and live traffic is still 1,500 a second, only 1,500 a second of spare capacity clears the backlog: another 2 hours. The service is "back" at 2 hours but not caught up until about 4. Say whether that meets the tolerance as the firm defines it, and what you would change (more submission capacity, or prioritising older payments) if it does not.

Detailed design. Pick the ledger write and the scheme submission. A common answer is an outbox: the debit and an "outgoing payment" record commit in one database transaction, and a separate worker reads the outbox and submits. That way a crash between debit and submission cannot lose a payment, and the worker can retry safely because the scheme request carries the same idempotency key.

UK GDPR: breach handling is a requirement

UK GDPR Article 33 requires a controller to notify a personal data breach without undue delay and, where feasible, not later than 72 hours after becoming aware of it, unless it is unlikely to result in a risk to people's rights and freedoms, and to document every breach. For a design, that is a non-functional requirement: you must be able to say quickly whose data was involved.

Put a number on it. If the audit logs are 40 TB and a full scan reads at 2 GB a second, one scan takes 40,000 / 2 = 20,000 seconds, about 5.6 hours. A few scans fit in 72 hours, with little room for investigation, so propose partitioning audit logs by customer and date and keeping an index from data item to customer.

Where it runs: one London region on AWS

Ask whether data must stay in the UK. The AWS regions list shows a single UK region, Europe (London), eu-west-2, with three availability zones. A UK-only design can spread across those zones, which protects against a data-centre failure but not a regional one. For geographic distance you need a second provider, the firm's own data centre, or the interviewer's agreement that a recovery copy in another European region, such as Ireland (eu-west-1), is acceptable. Each choice is a trade-off to state, not a default.

Building blocks UK interviewers probe

  • Idempotency and exactly-once effects. Message delivery is usually at least once; effects must be once.
  • Reconciliation. After an outage, how do you prove the ledger and the scheme agree? A daily or hourly reconciliation job is part of the design.
  • Third-party dependencies. Card schemes, payment schemes, identity checks and cloud providers. Draw them; operational resilience mapping exists to surface them.
  • Degraded modes. Can customers see balances while payments are queued? Saying what still works is often worth more than another replica.
  • Back-of-the-envelope figures. A day is 86,400 seconds; 1 million a day is about 12 a second; state your peak multiplier rather than assuming one.

Common UK-specific mistakes

  • Quoting "five nines" when the interviewer asked about an impact tolerance. Translate availability into minutes of disruption at peak.
  • Ignoring the payment scheme as a dependency. The service can be perfect and still breach tolerance when its downstream connection fails.
  • Treating UK GDPR as a legal footnote. Breach detection, audit logs and data maps are engineering work.
  • Assuming your cloud provider has a second UK region. Check: AWS, for example, lists one.

Tools for this job

Frequently asked questions

Do UK banks ask different system design questions?

The prompts look similar (payments, notifications, ledgers), but the follow-ups weigh failure more: how the service degrades, what is queued and replayed, how you reconcile and which third parties you depend on. Frame recovery in terms of an impact tolerance.

What is an impact tolerance in plain terms?

The longest disruption to an important business service a firm will accept before customers or markets suffer intolerable harm. The PRA requires it to include a time-based metric and to apply at peak times.

How does the 72-hour UK GDPR rule affect a design?

You need to detect a breach and work out whose data was involved fast enough to notify within 72 hours where feasible. That means audit logging, data maps and logs partitioned so a search takes hours, not days.

Should a UK design keep everything in the UK?

Only if the requirements say so; ask. If they do, and you design on AWS, its single London region gives three availability zones but no regional separation, so say how you would handle a full regional outage.

Sources and further reading

Open the System Design Interview Framework

More Interview Prep

Every tool in the Interview Prep section, free and in your browser.