systemdesigninterview.org

Defend the architecture when a dependency fails.

Requirements, load, failure and cost. Make the trade-offs explicit.

Sponsored bySocratify AI
Try a design decision

Peak load. One missing zone.

12,000 jobs/second · authored exercise

Zone ARunning
Zone BRunning
Zone CUnavailable
Test your fleet

Five decisions worth defending

01 Size for the failure state

Amazon asks candidates to validate the requirements: “Your interviewer will ask questions related to your design, and you should ask questions to complete and validate your design.” Atlassian makes the constraints concrete: “how you’d talk through constraints like reliability and cost, who you might partner with for help, and which technologies you’d use.”

“the servers in the remaining Availability Zones could carry the load.”

Mike Furr, Author, Amazon Builders’ Library · AWS Builder Center, Dec 5, 2019

Compare peak demand with surviving capacity, not the normal fleet.

  • Amazon Amazon
  • Atlassian Atlassian
System design · Original practiceCan this fleet survive a zone loss?Example

02 A lost reply is not a failed write

“A timeout or failure doesn't necessarily mean that side effects haven't happened.”

Marc Brooker, Author, Amazon Builders’ Library · AWS Builder Center, Dec 5, 2019

Choose one operation identity across repeated attempts.

Brooker’s retry rule is to “retry at a single point in the stack.” A timeout policy also needs a total deadline, an attempt limit and evidence that retrying can help.

  • Marc Brooker AWS
System design · Original practiceShould you retry the timed-out webhook?Example

03 One customer must not consume every worker

“reduces the blast radius by limiting infra-level incidents to a shard-level impact.”

Florencia Bonansea, Senior Principal Engineer, Atlassian · Inside Atlassian, Dec 15, 2025

Name both the isolated failure and the dependencies still shared.

In his EC2 account, van der Merwe protects the existing data plane: “VMs that are already running need to keep working.” A new configuration request must not become a dependency for every live request.

  • Zak van der Merwe AWS
System design · Original practiceOne tenant fills the queue. What changes?Example

04 Fresh enough is a business rule

“preventing race conditions or out-of-order events that could expose unauthorized data.”

Florencia Bonansea, Senior Principal Engineer, Atlassian · Inside Atlassian, Dec 15, 2025

Authorize the object version you will actually serve.

Meyer says an agent needs more than a ticket: “It needs the requirements, the relevant architecture, the decision history, and the constraints the team already knows.” Include the consistency requirement before accepting its proposed cache.

  • Dave Meyer Atlassian
System design · Original practiceCan you serve this cached issue?Example

05 Keep the delivery path independent

“When code is cheap, the bottleneck moves to judgment”

Zak van der Merwe, EC2 and DSQL control-plane engineer, AWS · All Things Distributed, Aug 4, 2026

Own what happens after the diagram becomes a running service.

Salanki’s team uses AI coding tools while trying “not to lose grip of architecture, how everything actually works.”

  • Peter Salanki CoreWeave
System design · Original practiceThe AI provider is slow. Does dispatch stop?Example

Sources

  1. Amazon, SDE III candidate guidance, at the time. Amazon Jobs, “SDE III/Sr. SDE Interview Prep”, Undated. Open the source. Checked against the published text.
  2. Atlassian, Engineering candidate guidance, at the time. Atlassian, “Atlassian engineering interview handbook”, Undated. Open the source. Checked against the published text.
  3. Mike Furr, Author, Amazon Builders’ Library, at the time. AWS Builder Center, “Static stability using Availability Zones”, Dec 5, 2019. Open the source. Checked against the published text.
  4. Marc Brooker, Author, Amazon Builders’ Library, at the time. AWS Builder Center, “Timeouts, retries, and backoff with jitter”, Dec 5, 2019. Open the source. Checked against the published text.
  5. Florencia Bonansea, Senior Principal Engineer, Atlassian, at the time. Inside Atlassian, “How We Unlocked Performance at Scale with Jira Platform”, Dec 15, 2025. Open the source. Checked against the published text.
  6. Zak van der Merwe, EC2 and DSQL control-plane engineer, AWS, at the time. All Things Distributed, “On building scalable control planes”, Aug 4, 2026. Open the source. Checked against the published text.
  7. Dave Meyer, Author, Inside Atlassian, at the time. Inside Atlassian, “How we’re evolving Jira for AI-native software development”, Jul 15, 2026. Open the source. Checked against the published text.
  8. Peter Salanki, CTO and co-founder, CoreWeave, at the time. The Stack Overflow Podcast, “Do you have what it takes to run AI in production?”, May 26, 2026. Open the source. Checked against the published text.

Last reviewed Nov 10, 2026