Use Template

Opens this plan in Hirezen, where one click makes it a position.

Cloud Architect interview questionsMigration review — the data centre exit round

A 60 min interview plan with a time-boxed script, what each question is for, and the signals to score against. Key skills: Cloud migration planning for a data centre exit: finding what must move together and what a wave plan cannot survive, designing the network for the months both sides run, and choosing managed or self-run PostgreSQL on a three-year cost that includes people..

The wave plan

22 min
What this section is for

Purpose

Runs over a migration pack sent 24 hours ahead. The company's product already runs in the cloud; its back office runs in a colocation data centre whose contract ends in 14 months and will not be renewed. The pack has four parts, and each of the plants below has to be in it as written. An inventory and draft wave plan: 31 low-traffic internal applications marked "rehost as VMs, wave 1"; the finance system's batch server marked "wave 1" while its ERP database, 600 GB, is "wave 3"; a payments file service marked "wave 2, two-week cutover"; a document-management system marked "rehost, wave 2", whose licence names the physical hosts it may run on; a 40 TB file server that the plan calls its largest risk and proposes to move by shipping an offline transfer appliance with a six-week lead time; and the customer portal's PostgreSQL 14 cluster, 1.1 TB, marked "move to the provider's managed PostgreSQL", which depends on an in-house C extension for tenant-aware search ranking whose author left last year. A dependency map from 30 days of flow data, showing that the finance batch sends about 9 million queries a night to the ERP database at 0.4 ms round trip, inside a six-hour window it currently finishes in 1 h 50 min. A network sheet: one 1 Gbps link to the cloud carrying a site-to-site VPN at 25 ms round trip; the data centre announces 10.0.0.0/8 as a single route, with servers in 10.0.0.0/16, storage in 10.1.0.0/16 and finance in 10.40.0.0/16; the cloud networks were created from defaults as 10.0.0.0/16 for production, 10.1.0.0/16 for staging and 10.2.0.0/16 for shared services; the draft plan says "route 10.0.0.0/8 over the VPN"; 14 inventory rows note IP addresses written into configuration; and the partner bank that exchanges payment files over SFTP allows only the data centre's public address range, which belongs to the colocation provider, and changes its allow-list only in the first week of a quarter, on eight weeks' notice. A cost sheet for the portal database: the managed service at $3,900 a month for a highly available pair, self-run on two VMs with storage and backups at $2,300 a month, the team's estimate that running it themselves takes 0.3 of an engineer at a loaded $180,000 a year, and a note that nobody has estimated removing the extension. Rehosting the 31 internal applications unchanged is deliberately correct. The file server is the red herring: over the existing link 40 TB takes under four days at full rate and about a week at half, so the appliance solves a problem the link does not have. Book 70 minutes; the closing exchange runs past the 60.

I'm [YOUR_NAME] and I lead the move out of our data centre at [COMPANY_NAME]. The pack is the plan as it stands today. Nothing has moved yet, so everything in it is still cheap to change — which is why you are seeing it now.

What this section is for

Purpose

Sets the round as a review of a plan that can still change, so the candidate's job is to find what it gets wrong rather than to present a migration methodology.

Nobody in this room wrote the plan or is attached to it. Tell me what is wrong with it as plainly as you would if you had inherited it on Monday.

What this section is for

Purpose

Removes politeness as a reason to soften findings. A candidate who still hedges every point here will hedge in a steering meeting too.

Here is the inventory and the draft wave plan. Tell me what has to move together, what cannot move as planned, and what you would check this week before anything moves at all.

What this question is for, and what to listen for

Purpose

The load-bearing question. Most candidates can describe rehost, replatform and refactor; the separator is whether they read the dependency map and the outside lead times before the labels, and whether they do the arithmetic that turns "wave 1" into a batch that cannot finish.

Signals to score

  • Reads the dependency map before the inventory's labels, and groups workloads by how they talk rather than by team or size
  • Finds the finance batch split from its database and works out the effect: 9 million queries at 25 ms is about 62 hours against a six-hour window
  • Makes the batch server and the ERP database one move group, or names the change to the batch that would let them move apart
  • Finds the bank's allow-list rule and works backwards from a quarter's first week to the date the payments cutover can happen
  • Asks who owns the data centre's public address range, and keeps payment files leaving through it until the bank has changed its list
  • Flags the licence tied to named physical hosts as something to settle with the vendor this week, before the wave that depends on it is booked
  • Agrees with rehosting the 31 internal applications unchanged, and says why a fixed deadline makes that right
  • Does the arithmetic on the file server and drops the appliance
  • Orders the work by lead time, starting the outside conversations first even where the move itself comes later
  • Names what would make them ask for the data centre contract to be extended, and by when they would need to know

Follow-up questions

  • The batch server moves in wave 1 and its database in wave 3. What happens on the first night?
  • The bank changes its allow-list once a quarter, on eight weeks' notice. When does that conversation have to start?
  • How long does 40 TB take over the link you already have?
  • Which workloads would you move without changing anything, and why?
  • What would make you ask for more time on the contract, and when would you have to know?

Fourteen months of two networks

20 min
What this section is for

Purpose

The coexistence network is the longest-lived temporary thing in any migration, and the pack's is broken before it is built: the cloud's default ranges overlap the data centre's, and the draft plan routes all of 10.0.0.0/8. This section reads how the candidate draws a network boundary that has to work for fourteen months and then disappear.

The draft plan says route 10.0.0.0/8 over the VPN. Design the network between the data centre and the cloud for the fourteen months both exist — addresses, routing, name resolution and the link itself.

What this question is for, and what to listen for

Purpose

Tests whether the candidate sees the address overlap before designing on top of it, chooses what to renumber by what is being kept, and treats names, the link and the end state as part of the boundary.

Signals to score

  • Sees that cloud production and staging overlap the data centre's server and storage ranges, so the draft route cannot work
  • Rejects renumbering the data centre, which is being emptied, and is cautious about renumbering cloud production during the migration
  • Proposes new cloud networks for migrated workloads from a range that overlaps nothing, taken from one address plan for the whole organisation
  • Limits what cloud production reaches in the data centre to a short list of flows, uses translation only for those if unavoidable, and says what translation breaks
  • Advertises specific routes instead of 10.0.0.0/8, so each side reaches only what it needs
  • Designs name resolution in both directions so migrated systems find each other by name, and treats the 14 rows with written-in addresses as a list to fix
  • Checks the 1 Gbps link against migration copies and coexistence traffic, and asks the lead time for a dedicated connection
  • Treats the link as a single point of failure for the finance move group and the payment files while both sides run, and says what happens when it drops
  • Describes the network on the day the data centre closes, so nothing temporary stays by accident

Follow-up questions

  • Cloud production is 10.0.0.0/16, and so are the data centre's servers. Where does a packet for 10.0.3.20 go?
  • Which side would you renumber, if either?
  • A migrated application has 10.0.1.14 written in its configuration. What happens to it after the move?
  • The link drops for an hour during wave 2. What stops?
  • What is left of this design the day after the data centre closes?

The database with an extension

18 min
What this section is for

Purpose

A managed-or-self-run decision the pack makes hard on purpose: the managed service cannot run the database as it is, the major version is about to lose community support, and the one number that would decide it has never been estimated.

The plan moves the portal database to the provider's managed PostgreSQL. Using the cost sheet and what you know about the extension, make the call — managed or self-run — and price it over three years.

What this question is for, and what to listen for

Purpose

Tests whether the candidate prices people as well as instances, turns an unknown into a break-even instead of a guess, and notices that the version, not only the extension, is part of the decision.

Signals to score

  • Notices that the managed service loads only extensions it supports, so an in-house C extension blocks the plan as written
  • Notes that PostgreSQL 14 leaves community support in November 2026, so either path includes a major-version upgrade, and asks what it costs on each
  • Prices both options over three years with people included: about $140,400 managed before any extension work, about $244,800 self-run
  • Turns the unknown into a break-even: removing the extension can cost up to about $104,000 — roughly seven engineer-months at the sheet's rate — before self-run is cheaper
  • Questions the 0.3 of an engineer, given that the extension's author has left and nobody is named as running PostgreSQL today
  • Proposes a time-boxed estimate of removing the extension before deciding, with a date that still fits the contract
  • Considers moving self-run first to meet the deadline and to managed later, and prices doing the move twice
  • Asks what the extension does for customers, since replacing the ranking may be cheaper than preserving it
  • States the decision and the evidence that would reverse it

Follow-up questions

  • Will the managed service load this extension?
  • PostgreSQL 14 leaves community support in November 2026. Which option does that hurt more?
  • How much could removing the extension cost before self-run wins?
  • The extension's author has left. What does that do to the 0.3 of an engineer?
  • If the rewrite cannot be estimated for six weeks, what do you do on Monday?

We're at time. What would you like to know about how this migration is actually going? Ask anything — who wrote this plan, what the contract allows if we slip, which team is least keen to move.

What this section is for

Purpose

A candidate who has run a migration asks about the contract, the owners and the dependencies nobody has mapped; one who has only planned one asks which migration tooling we use. The topics offered make their choice the signal.

And one true thing about where we are: [name one real obstacle in your own estate — a system nobody can explain, an address plan nobody owns, a deadline set before anyone counted the dependencies]. It would be on your desk in the first week.

What this section is for

Purpose

A specific, unflattering fact is what a candidate who does this work wants to know, and it discourages one who expected the hard part to be finished. Check that it is still true before you say it.

Cloud Architect interviews — common questions

Who is this Cloud Architect interview plan for?
It is written for the interviewer, not the candidate: the hiring manager, engineer or panel member running the Migration review — the data centre exit round for a Cloud Architect role. It gives you a 60 min script to follow in the conversation — 3 questions with what each one is for and the signals to score against — so you are not writing the round from scratch the night before.
What does the Migration review — the data centre exit round assess?
This round is focused on: Cloud migration planning for a data centre exit: finding what must move together and what a wave plan cannot survive, designing the network for the months both sides run, and choosing managed or self-run PostgreSQL on a three-year cost that includes people.. It works through The wave plan, Fourteen months of two networks and The database with an extension, scoring against 28 observable signals, with follow-up prompts on all 3 questions for going deeper where an answer is thin.
How is the 60 min split up?
The wave plan (22 min), Fourteen months of two networks (20 min), The database with an extension (18 min). The timings are there so the round stays on schedule and every candidate gets the same shape of interview — which is what makes two candidates comparable afterwards.
What other rounds should I run for a Cloud Architect?

A single round does not cover a whole role. The other rounds in this library for a Cloud Architect: