Skip to content
Tech Interview Prep
Technical interview guide

High Availability & Disaster Recovery

Designing for component failure as the expected case, and the RTO/RPO trade-off that shapes disaster-recovery strategy.

Read
26 min
Practice MCQs
25
Interview Q&A
25
Edition
v6
Editorial status
Reviewed

Scope: AWS Well-Architected reliability and DR guidance, Google Cloud DR architecture, and Azure reliability guidance accessed 2026-08-30..

Curated: · Written: · Reviewed:

25 questions

MCQ-3A balance API returns HTTP 200 with a well-formed body on every request, but a bad deploy serves yesterday's balance for 40% of accounts. Is the service available?
MCQ-4Why can three replicas in one availability zone still be a single point of failure?
MCQ-19A team runs its primary stack in one cloud account and its recovery stack in a second account with its own root credentials. Which benefit does that separate boundary provide?
MCQ-24A service runs one replica per availability zone. The zones share no power, network or hosts, and each replica has its own database. The team ships one build to all zones in a single wave. Which risk does this leave unmitigated?
MCQ-8A payments team runs a primary in us-east-1 that streams its write-ahead log to an asynchronous replica in eu-west-1. A failover drill to eu-west-1 leaves the recovery site missing the last 4 seconds of committed transactions. Which mechanism produces that gap?
MCQ-9A store keeps three replicas: one in us-east-1 and two in eu-west-1. Every write must be acknowledged by a majority (two of three) before the client's call returns, and the client always sends writes to the us-east-1 replica. The round-trip time between the regions is 45 ms, and local processing is negligible. Which pair of outcomes should the team expect?
MCQ-2A team runs a Postgres primary with a synchronous standby, so a commit is acknowledged only after the standby has the record. Failover to the standby is scripted and takes 45 minutes end to end. The business requires no more than 5 minutes of transaction loss and no more than 30 minutes of downtime. Which conclusion follows?
MCQ-1A payments API runs active-passive across two regions. Its DR plan sets RTO 4 h and RPO 15 min. At 09:00 the primary region loses power; the incident commander declares a disaster at 10:00; the standby region starts serving traffic at 13:30, using data whose last replicated write was at 08:50. Which statement about the plan's objectives is correct?
MCQ-22A checkout service runs a failover exercise. The team defines RTO as the time from failure injection to the checkout success rate returning to its pre-exercise baseline, and RPO as the gap between the last write confirmed on the standby and the failure. Targets are RTO ≤ 30 min and RPO ≤ 5 min. The exercise log: - 13:04:12. Last order confirmed replicated to the standby - 13:10:00. Primary terminated; paging alert fires - 13:19:00. Incident commander declares failover - 13:34:00. Standby promoted and accepting writes - 13:41:00. Checkout success rate back to the 99.9% baseline - 13:52:00. Exercise ends Which statement does the log support?
MCQ-20An active-passive service fails over by repointing the A record for api.example.com (TTL 30) from the primary 203.0.113.10 to the standby 203.0.113.20. The primary fails at 14:00:00 and the record is updated at 14:00:05. At 14:00:20 a monitoring probe that opens a fresh connection and sends a fresh query still reaches the primary. The probe's resolver fetched the record at 13:59:55. What explains this?
MCQ-7A checkout service calls an inventory service through a circuit breaker configured with a count-based sliding window of 20 calls, a failure-rate threshold of 50%, and a minimum of 10 calls in the window before the failure rate is evaluated. The breaker is closed. The last 8 calls to the inventory service, in order, were: failure, failure, failure, failure, failure, success, success, success. Call 9 now arrives. What does the breaker do with call 9?
MCQ-14A user-profile service runs active-active in two regions, us-east and eu-west. Each region accepts writes and replicates asynchronously to the other. Concurrent writes are resolved with last-write-wins on the whole record value, using the writer's wall-clock timestamp. A profile reads {name: "Ana", city: "Lisbon"} in both regions. us-east then writes {name: "Ana", city: "Porto"}; eu-west writes {name: "Anna", city: "Lisbon"} carrying a later timestamp. Both writes replicate and replication settles. What does the profile read in both regions?
MCQ-17A primary holds a 10-second lease that it renews every 3 seconds, and each write it sends to the shared storage layer carries a fencing token. The storage layer accepts a write only if its token is at least the highest token it has already accepted. The primary is paused for 12 seconds by a stop-the-world garbage collection. While it is paused, the coordinator promotes a replica, grants it fencing token 42, and the new primary writes successfully. The old primary then resumes and sends a write carrying token 41. What happens to that write?
MCQ-10An orders service runs PostgreSQL 15 on a primary with one asynchronous streaming replica (steady-state lag 200 ms). Backups are a nightly pg_basebackup at 02:00 plus continuous WAL archiving to object storage with archive_timeout = 60 s. At 14:02:07 an engineer runs DELETE FROM orders with no WHERE clause, removing 1.2M rows. The team must restore those rows with the least possible loss of other writes. Which recovery path meets that requirement?
MCQ-11A payments service runs PostgreSQL 15 on a primary with a nightly `pg_basebackup` at 01:00 to object storage and continuous WAL archiving. Its stated objectives are RPO 5 minutes and RTO 45 minutes. In a drill on Tuesday, the team declares the primary lost at 14:00:00, restores the most recent base backup, and replays every WAL segment present in the archive. The instance accepts connections at 14:35:00 and the smoke test passes. `orders` holds 1,204,880 rows, against 1,247,310 on the primary at 14:00:00, and the newest `orders.created_at` in the restored instance is 01:04:12. What does the drill show?
MCQ-25A payments service runs PostgreSQL 16 with asynchronous streaming replication from a primary in eu-west-1 to a standby in eu-central-1. The service returns HTTP 201 for POST /orders only after the database commit succeeds. The eu-west-1 region fails and the standby is promoted. The API gateway's access log shows 1,204 HTTP 201 responses for POST /orders in the final 6 seconds before the failure. On the new primary, the `orders` table has no rows created in that 6-second window, `pg_is_in_recovery()` returns false, and the connection pooler's health check returns 200. What does this evidence establish about the promoted standby?
MCQ-12A team must keep a 12 TB analytics warehouse recoverable after the loss of an entire region. The business accepts a 24-hour RTO and a 24-hour RPO, and a cost-governance policy blocks any always-on compute in the DR region outside a declared disaster. Which strategy should the team adopt?
MCQ-13An order API runs 40 instances in us-east-1. Its DR site in us-west-2 holds a Postgres read replica (lag under 3 s) and a launch template with the current AMI, but no running application instances. The last failover drill took 31 minutes: 14 launching instances, 7 running configuration management, 3 warming caches, 5 waiting out the DNS TTL, and 2 for manual approval. The RTO target is 15 minutes to serve production traffic, and the team may run up to 30% of production compute continuously in us-west-2. Which change is most likely to meet the RTO within that budget?
MCQ-15An active-passive service has a 30-minute RTO. The passive region already runs 40 app servers and 12 database replicas. Failover creates network peering, identity roles, and quota increases through the primary region's control plane. During a regional incident, that control plane's p99 rises from 200 ms to 8 s, and quota requests queue for 20 minutes. The runbook makes 300 control-plane calls sequentially, each retried up to 3 times. Which conclusion best explains why failover misses the RTO?
MCQ-6A service calls a dependency through a fixed pool of 40 connections. Healthy dependency latency is 5 ms, so the pool sustains about 8,000 requests/s. During an incident, latency rises to 2 s and every call hits the client's 500 ms deadline; the client retries each timed-out call twice more, immediately and with no backoff. Roughly how many logical requests per second can the service now sustain through the pool?
MCQ-5A payments API runs 30 instances in each of three availability zones. At peak it serves 90,000 rps, and one instance handles 1,000 rps at the 60% CPU target that keeps p99 under 200 ms. The autoscaler and its metrics store run in zone A, capped at 30 instances per zone. The objective: serve the full 90,000 rps peak through the loss of any single zone, holding p99 under 200 ms, with no manual intervention. Which change meets it?
MCQ-23A payments team's DR runbook targets a 4-hour RTO. Their quarterly exercise restores a 5 GB sample of the 4 TB production database and finishes in 22 minutes, so it passes. Running the same runbook against a full 4 TB copy, restore takes 3 h 10 m; the 90-minute write backlog then drains in a further 2 h 40 m; and the cache-warming job's 40k reads/s saturate the primary's 20-connection pool, pushing p99 from 40 ms to 180 ms. What does the full-scale run establish?
MCQ-16A payments service runs a single-writer PostgreSQL primary in eu-west-1 with a synchronous standby in eu-central-1. Leader leases live in a 3-node etcd cluster: two nodes in eu-west-1, one in eu-central-1. At T+0 a WAN partition isolates the two regions. The runbook says 'if the primary is unreachable for 30 s, promote the standby', and at T+45 s the on-call engineer runs `pg_ctl promote` on the eu-central-1 standby: a local postmaster signal that does not consult etcd. What is the outcome?
MCQ-18An active/standby service runs its primary in us-east-1 behind a warm standby in us-west-2. Writes require a lease held in a 3-voter store with one voter in each of us-east-1, us-west-2 and eu-west-1; the lease TTL is 10 s and the primary renews every 3 s, refusing writes as soon as it cannot confirm the lease with the store's majority. The store grants the standby a lease only after the primary's current lease has expired, and the failover controller sits in us-west-2 reading the store's log. Write unavailability is budgeted at 15 s per incident. At 03:11:58 a partition isolates us-east-1 from the other two regions. The log shows the last committed renewal at 03:11:56 and no lease granted since; the primary's last accepted write was at 03:11:58, its 03:11:59 renewal failed, and the standby holds every acknowledged write. At 03:12:09 the primary is still unreachable from us-west-2 and eu-west-1, and a health check running inside us-east-1 still reports it healthy. What should the controller do?
MCQ-21A single-writer service replicates asynchronously from primary A (us-east-1) to DR B (us-west-2) with p99 replication lag of 1.8 s. A fails; B is promoted 4 minutes later and takes writes for 6 h 20 min: 4.2M writes, of which 900k create keys A has never seen. A is then restored with its data intact as of the failure. The runbook says: repoint the application to A, then restart the original A→B replication stream. What is the result?