Cloud Architect interview questionsCost review — last month's bill round

Purpose

Separates a cost review from a savings checklist before it starts, so a candidate who prepared a list of generic optimisations knows the list is not the answer.

Purpose

Makes asking for data an ordinary move. What the candidate asks for — flow data, owners, the commitment's terms — is itself read in the first question.

Purpose

The load-bearing question. Nearly everyone finds savings; the separator is whether the candidate ranks charges by the design that produces them, and follows the NAT line past the route into the job that sends the traffic.

Signals to score

• Reads the flow-log page before ranking anything, and ranks by the design behind a charge rather than by the size of the line
• Traces the NAT charge to the usage-report job and names both causes: its route to object storage passes through the NAT, and it re-reads 31 days of events every night
• Proposes a private route to object storage that bypasses the NAT, and checks what that route costs on this provider instead of assuming it is free
• Proposes reading one new day of events a night instead of 31, and says what that does to the 281 TB
• Leaves the database alone and says why: 58% at peak and already reserved
• Splits the zone traffic into replication that buys resilience and fetches and calls that buy nothing
• Groups the unattached volumes, the snapshots with no expiry and the debug logging as housekeeping with a named owner, and does not lead with them
• Treats the commitment as a decision to manage for 22 more months rather than a line to fix, and asks what kind of commitment it is
• Puts a monthly figure on each change and says which figures are measured and which are estimates
• Asks who owns the job, the cluster and the commitment before handing out work

Follow-up questions

• The biggest line is the database. Why are you not starting there?
• 281 TB went through the NAT to a service in the same region. What is that traffic, and why is there so much of it?
• Suppose the route to object storage stopped going through the NAT tomorrow. What would still be wrong with that job?
• Which of these would you hand to someone else this week, and which would you keep?
• How much of the $9,600 between zones would you go on paying, and why?

Purpose

Tests whether the candidate knows what a zone boundary buys, can find the spend that buys nothing without giving up the spend that buys survival, and can say it to someone who does not care about brokers.

Signals to score

• Says what a single-zone cluster would stop surviving — the loss of that zone — and what stops when that happens: the usage events customers are billed on
• Separates the 190 TB of producer and replication traffic, which is the cost of zone resilience, from the 250 TB and 40 TB that buy nothing
• Proposes consumers reading from the replica in their own zone, and says what it asks of brokers and of clients
• Proposes routing service calls to their own zone first, and names what happens when one zone runs short of capacity
• Puts figures on it: about $5,800 a month removable, $3,800 a month kept
• Notes that a replica in the consumer's zone can be slightly behind the leader, and asks whether any consumer needs the newest records immediately
• Names where one zone is the right answer — development, staging, reprocessing that can be run again — and says why
• Asks what an hour without usage ingestion costs the business, so the $3,800 is weighed against something
• Answers finance in terms of what the money buys, not in broker vocabulary

Follow-up questions

• If that zone fails at ten on a Tuesday morning, what stops, and what is lost?
• Which of the 480 TB is the price of surviving a zone, and which is not?
• How does a consumer in one zone read from a replica in the same zone, and what does it give up?
• Is there anywhere in this estate you would accept a single zone?
• What sentence do you say to finance?

Purpose

Tests whether the candidate treats a discount as a bet on the future shape of the estate, uses the evidence of the last bet, and can say what they would sign rather than only what they would not.

Signals to score

• Says not before the known changes land, because pod requests at 3.1 times usage and the nightly job both inflate this month's compute
• Commits to the floor of expected usage after those changes, not to a share of this month's spend
• Works out the break-even: a commitment at discount d saves money only while more than 1 − d of it is used, and applies that to the existing one at 31%
• Asks whether the existing commitment can be exchanged, resold or applied to other work, and moves anything that can run on the previous family onto it
• Prefers commitment that follows the workload across instance families or regions, and says roughly what that flexibility costs in discount
• Checks the plans for the next three years for anything that moves compute — a second region, a new family, a managed service replacing a self-run one
• Proposes staggering commitments so they do not all end in the same month
• Says who signs, and what they are given to sign

Follow-up questions

• The last three-year commitment is 31% used. What went wrong, and what stops it happening again?
• What does 70% of this month's compute look like after the pods are right-sized?
• At a 40% discount, how much of the commitment has to be used before it saves anything?
• Is anything planned in the next three years that would move this compute?
• What would you sign next week, if anything?

Purpose

Tests whether the candidate makes cost an input to design decisions rather than an output someone reads a month later, and keeps the practice light enough that teams actually follow it.

Signals to score

• Asks for a data-flow drawing that marks every billable boundary a byte crosses — NAT, zone, region, internet — with the expected volume on each
• Wants a cost estimate in the design with its traffic assumptions written beside it, not only the price of the instances
• Proposes a unit cost the business recognises — per tenant, per million usage events — and follows it over time
• Gives every line of the bill an owner through account boundaries or tags, and says which is more reliable
• Alerts on unit cost or on one line's change rather than only on the monthly total
• Compares the estimate with the actual bill a month after launch, so estimates get better
• Keeps the review short enough that teams will do it without being chased

Follow-up questions

• Which design document would have predicted the NAT line?
• Tags are never complete. What do you do about that?
• What unit cost would you put in front of the leadership team, and who watches it?
• How would you know whether a design's cost estimate was right?

Purpose

A candidate who has lived with a bill asks who signs commitments and who owns the lines; one who has only optimised a bill asks which cost tool we use. The topics offered make their choice the signal.

Purpose

A specific, unflattering fact tells the candidate this round is looking for exactly what they would walk into, and warns off one who wants the spend already under control. Say it only if it is still true.