Every question on one page: open one, choose an answer, and see why at once. Signed in, your latest answer to each question counts, and you can answer again after 24 hours.
Curated: · Written: · Reviewed:
Reviewed 100Review pending 0
Answers to reviewed questions count toward this role and the overall leaderboard; your latest answer to each one is the one that counts.
MCQ-60Which of these is a legitimate reason to accept a single point of failure?⌄
MCQ-46A team runs a service across three availability zones in one region and is weighing whether a second region is worth the cost. They need to know exactly what the three-AZ layout already covers. Which statement about availability zones is accurate?⌄
MCQ-66A payments team's 99.9% availability SLO over a 28-day window allows about 40 minutes of downtime. Nine days before the window closes, incidents have used all 40 minutes. Product wants to ship a migration that forces two rolling restarts; SRE wants to freeze deploys. Both sides signed off on the SLO and its budget at the start of the quarter. What is the error budget doing in this standoff?⌄
MCQ-81An encrypted EBS volume holds the orders table, protected by a customer-managed KMS key whose policy grants the order service's instance profile use of that key. An attacker exploits an input-validation bug to run commands inside the service process, dumps the table using the service's own database credentials, and posts the rows to an external host. Which statement about the at-rest encryption is accurate for this incident?⌄
MCQ-5On PostgreSQL 16, session A runs at read committed:
```
BEGIN;
SELECT balance FROM accounts WHERE id = 42; -- 500
```
While A's transaction is still open, session B commits `UPDATE accounts SET balance = 300 WHERE id = 42;`. A re-runs the identical SELECT, gets 300, and commits. The team needs A's second read to return 500 again, and cannot change the query text. Which anomaly is this, and what prevents it?⌄
MCQ-26A per-customer running total must appear on every order row. Several orders can share the same ordered_at value, and the rule is that orders at the same instant count together: every row with a given ordered_at must show the same cumulative figure, covering all orders at that timestamp. Which expression produces that?⌄
MCQ-35A payments service replicates asynchronously to a cross-region standby. Its documented RPO is 5 minutes. During a nightly batch job, replica lag rises from 2 seconds to 4 minutes for 20 minutes, then recovers. The batch job is the only writer during that window. Which statement is correct?⌄
MCQ-62A payments team's DR runbook promotes a same-region, direct read replica that also serves the analytics dashboard. The stated RPO is 5 minutes. In a drill, promotion succeeded but 40 minutes of committed transactions were missing; the replica keeps up with the primary's write rate overnight when the dashboard is idle. Which is the most accurate reason the replica was not a valid failover target?⌄
MCQ-2A nightly batch job runs on EC2 instances in a private subnet. The subnet's route table sends 0.0.0.0/0 to a NAT gateway, the S3 bucket the job reads from is in the same Region and account, and the job pulls several terabytes a month. No S3 gateway endpoint has been added. What is the most significant consequence of leaving the route table as it is?⌄
MCQ-3An RDS MySQL instance sits in a public subnet whose route table sends 0.0.0.0/0 to the internet gateway, and it has a public IP. Its security group has a single inbound rule, TCP 3306 sourced from the application tier's security group, and an outbound rule allowing all traffic. The subnet's NACL allows 0.0.0.0/0 on all ports in both directions. After a Terraform apply, a host on the internet opens a MySQL connection to the instance. Which change in that apply explains it?⌄
MCQ-38A service runs on EC2 instances in a private subnet whose route table sends 0.0.0.0/0 to a NAT gateway. A compromised dependency now executes with the credentials attached to the instance profile, which can read a customer-data S3 bucket. Which control most directly limits how much of that data can leave the VPC?⌄
MCQ-58An EC2 instance in a private subnet runs `curl -v https://example.com` and hangs for about two minutes before failing with `Connection timed out`. The private subnet's route table contains 10.0.0.0/16 → local and 0.0.0.0/0 → nat-0abc123. The NAT gateway sits in a public subnet and its state is Available. The public subnet's route table contains only 10.0.0.0/16 → local. The instance's security group allows all outbound traffic and the subnet's NACL allows all inbound and outbound traffic. Which change makes the connection succeed?⌄
MCQ-43A team replaces a stored AWS access key with GitHub OIDC. The role's trust policy requires `token.actions.githubusercontent.com:aud` = `sts.amazonaws.com` and `token.actions.githubusercontent.com:sub` = `repo:acme/api:ref:refs/heads/main` (StringEquals). The deploy workflow is triggered by `on: push: tags: ['v*']`, the job has `permissions: id-token: write`, and it calls `sts:AssumeRoleWithWebIdentity`. A developer pushes tag `v2.1.0`. What does the assume-role call do?⌄
MCQ-18A team runs `terraform plan` on every pull request and `terraform apply` on merges to main. Both jobs assume the same CI service account, which holds administrator access to the production account. A reviewer proposes giving the plan job a read-only role and keeping write access only on the apply job. Which risk does that proposal remove?⌄
MCQ-8A Terraform plan against an unchanged configuration proposes to destroy and recreate a production RDS instance. The plan shows a destroy of `module.db.aws_db_instance.main` and a create of `module.rds.aws_db_instance.main`, with no attribute-level differences on the instance itself. Which explanation should you investigate first?⌄
MCQ-90A team's Terraform config names an S3 bucket `logs-${uuid()}`. They run `terraform apply`; it succeeds and creates `logs-4f2a`. With no edits to any file, they immediately run `terraform plan`, which reports `Plan: 1 to add, 0 to change, 0 to destroy` for a bucket named `logs-9c1b`. What explains the plan?⌄
MCQ-11A GET /invoices/{id} handler validates the caller's JWT signature and its invoices:read scope, then runs SELECT * FROM invoices WHERE id = ? with the path ID. What vulnerability remains?⌄
MCQ-57A service has 140 handlers. Each loads a record by the id in the request path, then calls `authorize(user, record)`. Two breaches in six months came from handlers where that second call was never written. The API gateway sees only the path and the JWT and has no access to record ownership; a lookup there would add a read to the hottest endpoints. Which change makes a handler that skips the call unable to return another account's record?⌄
MCQ-20A pipeline runs in private subnets and sends messages to Amazon SQS through a NAT gateway. A security review requires that a compromised workload cannot send data to any queue other than the four the pipeline owns. Which property of an interface VPC endpoint for SQS meets that requirement?⌄
MCQ-10A service calls a third-party API through a shared HTTP client with no timeout configured. The API accepts each connection and then takes 30 s to respond. What is the most likely failure?⌄
MCQ-28Retries are configured at three layers on one request path (the mobile client, the API gateway, and the service's HTTP client library) each set to three attempts when a call fails. During a partial outage the dependency begins failing a fraction of requests. What happens to load on that dependency?⌄
MCQ-45A checkout service calls a payment API from a fixed pool of 200 threads, each call with a 2 s timeout. At t=0 the payment API starts hanging, so every call holds its thread for the full 2 s. Requests arrive at 200/s. A circuit breaker is configured to open after 5 consecutive failures (a call counts as failed when it errors or times out) and stay open for 10 s. At t=1 s, how many of the 200 threads are occupied by payment calls?⌄
MCQ-51A checkout service shares one 200-thread pool across all outbound calls. The recommendations call has a 200 ms timeout and is returning in 150 ms with a 0.1% error rate, but it runs on every render and holds 180 of the 200 threads at peak, so checkout itself starts timing out. Which change keeps the recommendations slowdown from taking checkout down with it?⌄
MCQ-86A gateway fronts 10 services. Each service runs 4 single-threaded workers and takes 20 ms of worker time per request, so each service tops out near 200 rps. Rate limiting sits inside each service at 200 rps per client, and the limit is checked only after the request has been authenticated, which costs 5 ms of worker time. A client sends 20,000 rps spread evenly across the 10 services. What happens?⌄
MCQ-73An API gateway uses a token bucket limiter with capacity 100 tokens and a refill rate of 20 tokens per second. The bucket is full. At t=0 a client sends 150 requests in the same millisecond, then continues at a steady 10 requests per second. What does the limiter do?⌄
MCQ-74Your API rate-limits per key with a token bucket: capacity 10, refill 2 tokens/s. Rejected requests don't consume tokens. A client sends 30 requests at t=0; the 20 that don't fit get 429 with Retry-After: 1. The client waits 1 s, resends all remaining requests at once, and repeats until every request succeeds. When does the last request succeed?⌄
MCQ-77A gateway fans each request out to 20 independent shards in parallel and returns only after all 20 answer. Each shard has the same latency distribution: p50 5 ms, p99 50 ms. The gateway's own p99 is 160 ms. Which explanation accounts for the gateway's p99?⌄
MCQ-76Checkout calls the recommendation service behind the boolean kill switch `recs_v2` (no percentage rollout). During an incident the recommendation service's p99 rises from 90 ms to 3.5 s, checkout's 200-thread pool saturates, and its error rate hits 12%. On-call flips `recs_v2` off; the SDK polls every 30 s, and 10 minutes later the debug endpoint on three pods reports `recs_v2 = false`. Checkout's outbound request rate to the recommendation service is still ~1,800 rps and the pool is still saturated. The recommendation client is configured with max attempts = 1. What is the most likely cause?⌄
MCQ-53An ALB in front of a Java service returns a burst of 504s during a 10-minute traffic spike. Over that window CloudWatch reports HealthyHostCount at the full target count, UnHealthyHostCount 0, HTTPCode_Target_5XX_Count 0, and the 504s counted in HTTPCode_ELB_5XX_Count. Which reading do these numbers support?⌄
MCQ-59A customer reports that roughly 1 in 200 requests to your API returns a 500, but the service's error-rate dashboard sits flat at 0.005% and p99 latency is unchanged. The service runs 40 instances across 3 zones; this customer sends 1% of total traffic. Every request is already tagged with tenant, serving instance, zone, and client version. What is the most useful first step?⌄
MCQ-64A report groups orders by the region of the customer who placed each order. The orders table holds 48,200 rows; the report returns 4 regions whose counts sum to 44,900. The region column has no NULLs, and the query has no WHERE, HAVING or LIMIT clause. Which cause explains the shortfall?⌄
MCQ-15An events table holds 200M rows across 50,000 tenants and 200 distinct status values, both roughly uniform, so one (tenant_id, status) pair matches about 20 rows. The hot query is `SELECT ... WHERE tenant_id = ? AND status = ? ORDER BY created_at DESC LIMIT 20`, and it is issued with every status value. Which index should the team add?⌄
MCQ-47A service pages an `events` table in PostgreSQL with `ORDER BY created_at, id LIMIT 20 OFFSET ?`, backed by a composite index on `(created_at, id)`. `EXPLAIN (ANALYZE)` shows a single Index Scan node with no Sort, no Materialize and no temp files, and the plan is served from cache. Page 1 returns in 2 ms; page 10,000 (OFFSET 200000) takes 430 ms. What explains the growth?⌄
MCQ-19A `posts` table holds roughly 400 million rows and new rows are inserted at the head continuously. A client pages through it ordered by `created_at DESC`, and the team needs no duplicates or skipped rows across pages plus page 500 costing about the same as page 1. Which approach?⌄
MCQ-30A telemetry service writes 50,000 readings per second into a single table. Each reading carries a sensor ID (40,000 distinct sensors, all written at the same rate), a millisecond-resolution collection timestamp, and a tenant ID (three tenants, one of which produces 85% of all readings). Writes arrive continuously and are spread evenly across sensors and across time. Which partition key is most likely to create a hot partition?⌄
MCQ-4A read-through cache in front of a primary database absorbs 95% of reads and holds 40 million keys totalling about 2 TB; application instances have 4 GB of memory each. The origin serves about 5,000 queries per second in total, and the reads must be strongly consistent, so asynchronous replicas cannot serve them. Traffic requests roughly 40 million distinct keys per hour, each key about once, so concurrent misses are almost always for different keys. Every deploy flushes the cache and the origin saturates within seconds. Deploys ramp gradually, the new version takes 1% of traffic per hour over about four days, and this release changes the serialized format of the cached value, so new code cannot read entries written by the old version. Which change removes the cause of the outage?⌄
MCQ-33A Lambda function (Node.js 20 runtime) creates a `pg` Pool at module scope and reuses it across invocations. A load test drives 300 concurrent invocations against a PostgreSQL instance with `max_connections = 100`, and requests start failing with `FATAL: sorry, too many clients already`. Ten minutes after the load test ends, `pg_stat_activity` still shows 97 connections from the function's user, all in state `idle`. What explains the failure?⌄
MCQ-37Two teams own two services that share one Postgres schema on one instance. Each team ships schema changes on its own release train, and the two services need different constraints on the same tables. Which is the strongest argument against sharing the schema?⌄
MCQ-72A team has split a monolith into two services and is reviewing whether the boundary sits in the right place. Which observation most strongly indicates the boundary is wrong?⌄
MCQ-54A payments platform runs 60 services across 9 teams. Its Confluence architecture diagram was last edited 14 months ago, and 11 of the 23 service-to-service dependencies added since then are missing from it. The service catalog's dependency list is hand-entered by each team and was last reconciled 11 months ago. Last month an on-call engineer followed the diagram's call path during a Sev-1, paged the wrong team, and time-to-mitigate ran 47 minutes against a 12-minute median for that severity. The team runs about four Sev-1s a quarter and has no headcount for a recurring manual diagram chore. Which practice keeps the diagram accurate under these constraints?⌄
MCQ-79A payments team runs three regions and deploys config through a canary stage that covers region A only. Over 14 months it logged nine Sev-2 incidents; six shared one mechanism: a config change rolled to all three regions in a single step, so regions B and C took it unverified. Each postmortem was published within five days, named the canary gap as the mechanism, and closed with three to five actions, all detection-side ("alert when a config change reaches more than one region"). Every action had a named owner and a due date, and all were completed. Which change to the postmortem process would most likely have stopped the repeats?⌄
MCQ-98A payments team's on-call rotation pages about 45 times a week. A review of the last 300 pages found 190 were closed with no action taken, and most of the rest fired on host CPU above 80% rather than on failed or slow checkouts. Which change most improves the rotation's sustainability?⌄
MCQ-83A team proposes halving instance size because average CPU across the fleet sits at 15%. Peak memory utilization on the same fleet is 40%. Which check should come first?⌄
MCQ-31A platform runs a stateless API on 120 m5.xlarge instances across three AZs in us-east-1, 24/7, behind a 99.95% availability SLO. The architecture review has frozen the instance family at m5 for the next three years, the Graviton migration is deferred past the term, and no serverless or Fargate workloads are in scope. Load is expected to grow, so the team will resize within the m5 family (xlarge to 2xlarge) and may shift capacity between AZs. Which 3-year commitment gives the lowest total cost for this plan?⌄
MCQ-100A platform team's service is billed per request. Six weeks into last quarter the team shipped a caching layer, and 30% of the service's traffic was migrated to a sister team; total spend for the quarter fell 18%. Leadership wants evidence that the team's cost work changed how it builds, not that the bill moved for unrelated reasons. Which measurement gives that evidence?⌄
MCQ-49An 8-person platform team runs PostgreSQL 16 on three EC2 instances (one primary, two replicas) with Patroni for automated failover and pgBackRest for backups. Over the last six months, failovers completed in 38–52 s (median 44 s), all 26 weekly restore drills passed, and database operations consume 2 engineer-hours per month at a fully-loaded $110/hour. The dataset is 12 TB and grows 400 GB/month. The managed equivalent quotes $9,800/month; the current EC2, EBS and S3 bill is $3,100/month. Which decision is most defensible?⌄
MCQ-89A team is deciding whether to keep a Redis cache in front of a slow read endpoint or spend the sprint fixing the query. Which observation would justify removing the cache and fixing the query instead?⌄
MCQ-91An incident corrupted some user records, and you must both correlate one user's requests across four services and, once you find the failing request, identify the account so support can notify the affected person. The only stable identifier in the request path is the user's email address. Logs from all four services ship to a third-party aggregator and are readable by every engineer in the company. Which logging approach should you adopt?⌄
MCQ-44A multi-tenant reporting API caches responses in one shared Redis keyspace. The cache key is built from the HTTP method, the normalized path, the sorted query string and the negotiated response format. Tenant identity is resolved from the caller's subdomain and is not part of the key. Entries carry a 300-second TTL and are deleted by exact key whenever the underlying report is written. Within one TTL window, two different tenants request the same report path, with the same query string and the same negotiated format. Which failure does this key produce?⌄
MCQ-94A payments API runs 12 pods across 3 AZs (4 per AZ). At peak it serves 6,000 rps; each pod is load-tested to 800 rps at p99 180 ms, and the SLO is p99 < 300 ms. The team wants to validate that losing one AZ is safe. They have a tested abort that scales the deployment back to 12 pods and drains traffic in 90 seconds, and the on-call team is only available during business hours. Which experiment should they run first?⌄
MCQ-84A team is strangling a monolith. A facade now routes every profile-page read to a new profile service, which reads from a replica of the monolith's users table; replication lag is up to two seconds. Address updates still go to the monolith, and the facade routes them there unconditionally, with no session affinity. A user updates their address, gets a 200, and immediately reloads the profile page. What does the user see?⌄
MCQ-85A payments service runs 40 instances on Postgres. Deploys are rolling and take 12 minutes. The team's rollback runbook allows rolling back to the previous release for up to 36 hours after a deploy. A compliance deadline requires the `legacy_currency` column to be gone within 72 hours. The current release still reads and writes the column. The migration tool runs each migration once, at the start of a deploy, before the new code serves traffic. Which plan removes the column without breaking a rollback and meets the deadline?⌄
MCQ-32A 400-million-row `orders` table on PostgreSQL 15 must gain a `region_id` column that is NOT NULL and derived per row from a `regions` lookup. Deploys are rolling, so the old and new app versions serve traffic at the same time, and there is no maintenance window. Which sequence is safe?⌄
MCQ-80You are migrating a 50 TB MySQL 8.0 primary to a new store. The primary backs a third-party application you cannot modify, foreign keys span tables, and it must stay online throughout. Binlog retention is 14 days and the bulk load takes about five days. Which cutover plan fits?⌄
MCQ-87A payments API runs PostgreSQL 16 with `synchronous_commit = on` and `synchronous_standby_names = 'eu_standby'`. The primary is in us-east-1; the synchronous standby is being moved from us-east-1 (0.5 ms RTT) to eu-west-1 (78 ms RTT). Each write transaction issues 3 statements, does 4 ms of local work, then commits. The connection pool holds 32 connections and the API offers 400 write transactions per second. What happens after the move?⌄
MCQ-48A nightly batch job and the request path deadlock two or three times a night on the `items` table in PostgreSQL 16. The last deadlock graph shows the batch job holding a lock on row 8812 and waiting for row 4401, while a request transaction holds 4401 and waits for 8812. The batch job runs `UPDATE items SET ... WHERE id = ANY($1)`, where `$1` is built by a query with no `ORDER BY`; the request path updates the same rows in ascending id order. Which change most directly removes the deadlock?⌄
MCQ-6Checkout handler (Python 3.12), one attempt per HTTP request:
```python
def pay(order_id, amount):
key = str(uuid.uuid4()) # new key per attempt
charge = processor.charge(key, amount) # dedupes on key for 24 h
db.execute("INSERT INTO orders (id, charge_id) VALUES (%s, %s)",
order_id, charge.id)
return charge
```
The client times out after 10 s and retries the same order up to 3 times. During a processor slowdown the first call takes 40 s; the client retries twice more. All three calls reach the processor and each returns a distinct charge id. One order, three charges of $49.99.
Which change collapses the three attempts to one charge?⌄
MCQ-16A webhook receiver's handler writes to Postgres and has no other side effects. The provider retries any delivery that does not return 2xx within 5 seconds, always reusing the same `event_id`, and some events are being applied twice. The team needs each event to take effect exactly once, with no duplicates and no lost events. Which change achieves that?⌄
MCQ-39A payments service reads from a broker, calls a charge API that accepts an idempotency-key header, then commits its offset. A broker failover caused some customers to be charged twice. Which change removes the double charges while still charging every customer?⌄
MCQ-71A consumer reads `Debit` events from a queue with at-least-once delivery. For each event it runs `BEGIN; UPDATE accounts SET balance = balance - 100 WHERE id = 7; COMMIT;` and acks only after the commit. Account 7 holds 500. The consumer crashes after the commit but before the ack, and the broker redelivers the same event. The team's fix adds `INSERT INTO processed_messages(message_id) VALUES ($1)` as the first statement of that same transaction, with `message_id` as the primary key. After the redelivery is handled, what is account 7's balance, and why?⌄
MCQ-27A payments service updates a balance row and must publish a `BalanceDebited` event for that update. The event must never be published for an update that rolled back, must never be missing after an update commits, and consumers must see it within about a second. The broker is Kafka, which has no XA support. Which approach meets all three requirements?⌄
MCQ-17A checkout saga runs charge-card, then reserve-stock, then book-courier, each a local transaction with a compensating action. The payment provider supports authorization holds: a hold reserves funds without posting a charge, and it can be captured or released. During flash sales, failures are common at both the stock and courier steps, and support is getting complaints from customers who see a charge for an order that was never fulfilled. Which change removes the complaint at its source?⌄
MCQ-56A Kafka consumer group reads an orders topic with 12 partitions. Each record is processed synchronously; on any exception the consumer logs and retries the same record with exponential backoff, never committing the offset. A downstream schema change now makes 0.3% of records fail with a permanent 400 validation error. Over 20 minutes, lag on the affected partitions grows from 0 to 1.8M records while the other partitions stay near zero. What should the team change?⌄
MCQ-23A nightly bulk import for one tenant enqueues 200,000 jobs in a single burst onto the shared job queue, and those jobs succeed on first delivery. Other tenants' background jobs, enqueued after the burst, do not run until the import drains. The queue is a single FIFO served by a fixed pool of 20 workers. Which change most directly bounds how long the other tenants wait?⌄
MCQ-9A consumer fleet autoscales on queue depth and prefetches 100 messages per instance, so the broker takes a message out of the queue as soon as a consumer fetches it. Since 09:10, one downstream call has slowed to 90 seconds, so each consumer sits on a single message for that long while the rest of its prefetched batch waits. Which signal would have shown consumers falling behind first?⌄
MCQ-67A payments service moves its order-total recalculation into an AFTER INSERT row trigger on `orders`; the trigger updates a shared `inventory` row. The load test runs 200 concurrent connections, each inserting a single order row, and `orders` has exactly one trigger, which runs as the table owner with the same privileges as the application role. p99 insert latency rises from 12 ms to 90 ms and lock waits pile up on `inventory`. Which property of the trigger explains this?⌄
MCQ-12On PostgreSQL 16, a team runs `ALTER TABLE events ADD COLUMN ingested_at timestamptz NOT NULL DEFAULT clock_timestamp();` against a 500-million-row table during peak traffic. The primary risk is that:⌄
MCQ-24A dashboard issues reads continuously against a materialized view that is refreshed hourly with plain REFRESH MATERIALIZED VIEW; each refresh takes about 90 seconds. Which risk should be called out in review?⌄
MCQ-41A payments service runs blue-green deploys: two identical fleets behind one load balancer, both pointing at the same Postgres cluster. Release 42 renames the column `amount_cents` to `amount_minor`. The team wants to switch traffic back to blue within a minute if error rates spike. What must the release process guarantee for that rollback to work?⌄
MCQ-69A payments queue feeds a pool of 20 workers. Steady-state depth is about 200 messages, and the SLO is 5 minutes from enqueue to completion. Last month one worker hung retrying a malformed message for 40 minutes while the other 19 kept draining the queue. Depth stayed near 200 and the p99 message age stayed under a minute, so no alert fired, and that payment breached the SLO. Which alerting rule should the team add?⌄
MCQ-99A repository's GitHub Actions workflow is triggered by `pull_request_target`, checks out the pull request's head commit, and runs `npm ci` in a job whose environment holds the repository's `AWS_DEPLOY_ROLE` credentials. A fork PR from a contributor whose earlier PRs have already been merged adds a `postinstall` script to `package.json` that reads `process.env` and POSTs it to an external host. What is the outcome?⌄
MCQ-97A team manages a production RDS instance in Terraform, with nightly drift detection against the applied state. During a Sev-1 overnight, an engineer changed the instance class from db.r6g.large to db.r6g.2xlarge in the cloud console to absorb a load spike. The change is legitimate and must stay until quarter end. Drift detection now reports the instance as drifted. Which response reconciles the environment correctly?⌄
MCQ-50An internal billing system can tolerate 8 hours of downtime and at most 1 hour of data loss if its primary region is lost. The stack is a stateless application tier over a 400 GB PostgreSQL database, both rebuilt from infrastructure-as-code templates and machine images. Which disaster recovery strategy meets these objectives at the lowest cost?⌄
MCQ-7A payments-reporting service must survive the loss of its primary region with an RTO of 15 minutes and an RPO of 1 minute. The workload is a stateless application tier over a 2 TB PostgreSQL database that supports asynchronous cross-region streaming replication with steady-state lag under 5 seconds. The team's last DR test measured a 4-hour restore from a cross-region snapshot and a 40-minute cold launch of the application tier from AMIs and IaC. Finance has rejected any design that runs full production capacity in two regions. Which strategy satisfies both targets?⌄
MCQ-63A payments team's disaster recovery runbook targets a 4-hour RTO. It is reviewed quarterly and every step is documented, but it has never been executed end to end; the last regional outage was handled by improvising and took 11 hours. Staging mirrors production's instance types and topology but uses its own DNS zone, its own IAM roles, and a sandboxed payment provider. Which action best establishes whether the 4-hour objective is achievable?⌄
MCQ-36A profile service stores each user record in a 3-node Cassandra cluster (replication factor 3, single region). Writes use LOCAL_QUORUM, reads use LOCAL_ONE. A user saves a new shipping address; the client reloads 150 ms after the write is acknowledged and sees the old address, and a manual refresh 2 s later shows the new one. Replication lag is p50 40 ms, p99 400 ms. The team wants the smallest change that guarantees a client sees its own write on the next read, without changing the write path.⌄
MCQ-22A service replicates each key to three regions and uses quorum reads and writes (R=2, W=2). Monitoring shows no network partition in 90 days, yet p99 write latency is 12 ms when the quorum stays in-region and 140 ms when it must cross regions. Which statement is accurate about the trade-off the team faces today?⌄
MCQ-29Your checkout service carries a 99.9% availability SLO over a rolling 30 days, which allows 43 minutes of full outage per window. Last quarter, most pages did not correspond to a customer-visible incident, and both real incidents were reported by customers first. The team now wants the pager to fire only when the incident, left alone, would use up the month's allowed downtime before the next on-call handover in 12 hours, and to stay silent for slower degradation. Which condition should page?⌄
MCQ-34A checkout service defines its SLI as the fraction of requests that return a non-5xx response, measured over a rolling 28-day window. Its internal SLO is 99.9%. The customer SLA promises 99.5% availability with service credits. In the current 28-day window the service has handled 4,000,000 requests, of which 5,200 returned 5xx. Which statement about this window is correct?⌄
MCQ-88Postmortem of a 9-day intrusion in a 400-account AWS estate: an attacker chained an SSRF in a public-facing EC2 web app to reach the instance metadata service, pulled the instance role's credentials, and used them to read the customer-data S3 bucket. The role's sessions were capped at 6 hours, and the attacker re-fetched credentials through the same SSRF roughly every 5 hours. CloudTrail and GuardDuty were enabled in every account, the app subnet could reach the database subnet, and the bucket was not public. Which single change would have broken the chain at the step the attacker actually used?⌄
MCQ-93A 180-service platform runs Istio with STRICT mTLS mesh-wide: every workload gets its own SPIFFE identity, and all pod-to-pod traffic is encrypted and mutually authenticated. No AuthorizationPolicy resources are defined anywhere in the mesh. An attacker exploits an SSRF bug in the public-facing checkout service and gets code execution in its pod. Which outcome should the team expect?⌄
MCQ-52A platform team runs 300 services on Kubernetes. A security review found that a compromised pod can set its own labels and, on their CNI, reuse another pod's IP address, so any authorization decision based on those attributes can be forged. Which service mesh capability closes that gap?⌄
MCQ-70A service runs in staging and production as separate Kubernetes Deployments, each with its own service account. A staging compromise last quarter exposed production database credentials. The team's rule is that no long-lived credential may be introduced and a compromised staging workload must not be able to read production values. Which design satisfies both?⌄
MCQ-61One state file holds the shared VPC, IAM roles and 40 service stacks. Every apply is gated by a reviewed plan, and last month a service change that edited a shared security group rule passed review and took every service down for 40 minutes. Which change would most reduce the blast radius of the next such mistake?⌄
MCQ-95A company is adding a fourth VPC that must be peered with three existing VPCs and attached to the on-premises network through a transit gateway. The existing VPCs use 10.16.0.0/16, 10.17.0.0/16, and 10.18.0.0/16. The on-premises network advertises 10.0.0.0/12 and 192.168.0.0/16. Corporate IPAM policy requires every cloud VPC to be allocated from 10.0.0.0/8. The new VPC needs a /16 block. Which CIDR assignment satisfies all of these constraints?⌄
MCQ-96A SOC 2 Type II audit covers 1 January to 31 December, with fieldwork in January of the following year. The auditor samples 40 production changes: 22 from January to June and 18 from August to December. Approval records live in a table that retains 180 days, and the deploy service's role can UPDATE and DELETE rows in it. Which statement about the evidence the team can produce is correct?⌄
MCQ-75A team writes its audit trail into the same cloud account as the workload it records. In their last incident, the attacker used that account's admin role to stop the trail and delete its stored log objects, and for the next several hours no events were recorded at all. They are moving the trail to a dedicated logging account. Which property of that move is what closes the gap the incident exposed?⌄
MCQ-40An order-history endpoint runs a four-table join over orders, order_items, customers and addresses. p99 is 320 ms; EXPLAIN ANALYZE attributes 280 ms to the join and 40 ms to serialization. Every join key is indexed and the planner uses hash joins; a query rewrite and a covering index on order_items each moved p99 by less than 5 ms. Connection-pool wait is 0.4 ms at p99, and the primary runs at 30% CPU with no lock waits. The four tables take a combined 3 writes per minute, and requests are per-customer: the top 1% of customers generate 4% of requests. Which change does this evidence justify?⌄
MCQ-78A service catalog runs on PostgreSQL 16: 400,000 `services` rows and 2,100,000 `depends_on` edges, a 12 GB working set on a 64 GB instance, so it stays resident. A recursive CTE that walks `depends_on` from one service returns in 30 ms at depth 3; the same query at depth 6 takes 3.8 s. `EXPLAIN ANALYZE` shows an index scan on `depends_on(parent_id)` at every level, with row estimates within 10% of actual. Which conclusion does the evidence support?⌄
MCQ-68A payments service runs PostgreSQL 16 on one primary: 16 vCPU, 64 GB RAM, 300 GB of data with a 30 GB hot working set. At peak it takes 3,000 write TPS, CPU sits at 85%, disk queue depth stays under 1, and p99 commit latency has risen from 8 ms to 45 ms over the past month. Every transaction writes to three or four tables, reporting queries join across them, and reporting already runs against a dedicated replica. What should the team do first?⌄
MCQ-65A payments team runs a nightly 40-minute aggregation job against the streaming replica that also serves as the failover target for the primary, which carries a 5-minute recovery point objective (RPO). The job finishes inside its window, its results are correct, the analytics team already holds read credentials, and the dashboards it feeds tolerate data up to 15 minutes stale. The replica's volumes are already paid for and sit idle overnight. Which is the best reason to move the job to a separate analytical store?⌄
MCQ-92Checkout service p99 is 180 ms against a 150 ms SLO. Every request opens a DB connection to write the order, so the pool of 20 is used regardless of cache. A trace sample of the 200 slowest requests shows 178 spent 90-140 ms waiting for a pool connection; the other 22 were slow on a downstream payment call. Redis cache hit ratio is 92%, and cache-miss p99 latency is 120 ms. Which change is most likely to bring p99 under the SLO?⌄