Next chapterRead the article

Infrastructure8 min read

Moving a core banking system without the customers noticing

Racks of servers in a dark room, threaded with bundles of orange and green cables

A successful core migration looks like nothing at all. On Monday morning the cards work, the balances are right, and nobody outside the project knows anything happened. Here is what that quiet Monday actually costs.

The highest compliment a core banking migration can receive is silence. No news story, no queue at the branch, no surge of calls on Monday morning. The customers wake up, the cards work, the salaries arrive, and not one of them has any idea that the system at the centre of the bank was replaced while they slept.

Achieving that silence is among the hardest things a bank ever does. The core is the system of record. Everything else, the mobile app, the ATMs, the card switch, the reports the regulator expects, leans on it. Replacing it has been compared to changing the engine of an aircraft in flight, and the comparison flatters aircraft engines, which have fewer integrations.

We have worked on enough of these to have opinions. What follows is not a methodology, and it will not fit every institution. It is what we wish someone had told us the first time.

Most cutovers are lost long before the weekend

When a migration goes badly, the post-mortem usually points to the weekend: a script that failed, a reconciliation that would not balance. Read closer and the cause is almost always months older. A product nobody had documented. An interface everyone assumed had been retired. A decision that was never quite made, because making it would have been uncomfortable.

The weekend is simply when the bill arrives. So the most useful thing a project can do early on is to make bad news cheap. If reporting a problem in month three costs someone an awkward meeting, and reporting it in month nine costs them their reputation, people will wait until month nine. The tone a sponsor sets in the first steering committee has more effect on cutover weekend than any tool.

Know what you are really moving

Every bank believes it knows what its core does. Every bank is partly wrong. Twenty years of adjustments leave behind products that were closed to new customers but still have live accounts, fee rules with exceptions nobody remembers agreeing, and batch jobs that exist because a report once needed them.

The inventory nobody wants to do

Before any design work, list what actually touches the core. Not what the architecture diagram says, but what the logs show over a full quarter, including month end and year end. It is slow, tedious and invariably surprising. We have yet to see an inventory that did not turn up something important and forgotten.

Decide what is not coming

A migration is the one chance a bank gets to stop doing things. Every dormant product and unused interface you carry across has to be built, tested and supported on the new platform. Moving something because nobody is sure it can be turned off is the most expensive decision on the whole programme, and it is usually made by default.

  • Products with a handful of remaining accounts are often cheaper to close or convert than to migrate.
  • Interfaces with no traffic for a year should be switched off now, on the old system, and the phone left to ring.
  • Reports nobody can name a reader for can go the same way, with the same test.
A bundle of glowing blue optical fibres spreading out against a dark background
Every connection to the core is one more thing that has to work on Monday morning.

Rehearse until it is boring

There is one rule we would not bend on. The cutover should be rehearsed end to end, against a full copy of production data, until it is boring. Not a sample. Not a subset chosen because it migrates cleanly. All of it, on the same timetable, with the same people, including the ones who will be doing it at four in the morning.

The first rehearsal is always a shock. It runs far longer than planned and exposes data no one knew existed. That is what it is for. By the third, the team knows how long each step really takes. By the fifth, someone is bored, and that is the moment to start believing the plan.

If the rehearsal was exciting, you are not ready. The weekend itself should feel like the sixth time you have done it, because it is.

Operations

Each rehearsal also answers a question no document can: whether the job fits in the window. If the data takes forty hours to move and the bank can only stand still for thirty, no amount of good intentions will help. You find this out in rehearsal, or you find it out on Sunday night.

Reconciliation is the real test

A migration has not succeeded because the scripts finished. It has succeeded when every account, every balance, every accrued interest figure and every pending transaction matches on both sides, and someone in finance who was not on the project is prepared to sign that it does. Automate this early, run it after every rehearsal, and treat a difference of one cent as seriously as a difference of a million. Small differences are usually the visible edge of something larger.

The weekend itself

By the time it arrives, cutover weekend should be the least interesting part of the project. Everyone knows their step, what comes before it and who to call. The plan runs to the minute, and there is a person whose only job is to watch the clock.

Friday night

Channels are closed in a set order, with customers told in advance, in Arabic and English, exactly what will be unavailable and for how long. Final batches run on the old system. Then the old system is frozen, and from this point the two sides must not drift.

Saturday

Data moves, and reconciliation begins. This is the long, quiet stretch. The temptation at this stage is to fix small things as they appear, and it must be resisted. Anything changed outside the rehearsed plan is, by definition, untested.

Sunday

The interfaces come up one by one, and real test transactions go through real channels: a card at a real ATM, a transfer through the production app, using accounts belonging to staff who agreed to it beforehand. Then the decision is taken.

Deciding when to go back

Every plan has a point of no return, and the most important thing about it is that it is written down before anyone is tired. At some moment on Sunday, going back becomes more dangerous than going on. Before that moment, the bank needs a clear rule for what would make it roll back, and who has the authority to say so.

  1. Agree the rollback criteria months ahead, in plain numbers rather than judgement calls.
  2. Name one person who makes the decision, and a deputy.
  3. Rehearse the rollback as seriously as the migration, including how long it takes.
  4. Make sure nobody will be punished for calling it, because a team afraid to roll back will talk itself into going forward.

Nobody wants to use this plan. Having it, and having rehearsed it, is what lets people think clearly at three in the morning.

Curving highway interchanges seen from above at night, lit against the dark
On cutover night, traffic has to find its new route first time, from every direction at once.

Running two systems at once

Not every bank moves in one weekend. Moving customers in waves, starting with staff and then a small branch, lowers the risk of any single night, and many institutions rightly prefer it. It has its own price, though, which is often underestimated.

For the length of the programme the bank is running two cores, and everything between them has to work: a transfer from a customer on the old system to one on the new, a statement that spans the move, a report that has to add both together. Each wave is a smaller weekend that must be planned just as carefully. Long programmes also meet the enemy that big-bang projects avoid, which is fatigue. The team that was sharp in wave one is tired by wave six, and the business is asking why it cannot launch anything new.

There is no universally right answer. What matters is making the choice deliberately, with the true cost of both options on the table.

The morning after, and the month after

The first working day is when the things nobody could test finally happen: real volumes, real customers, the first end of day under load. Have the whole team on hand and visible, with engineers sitting next to operations rather than on a call. Watch the call centre as closely as the dashboards, because customers will usually tell you about a problem before your monitoring does.

A server rack lit green in a dark room, with cables looped along its side
Hypercare is mostly watching. Keep the people who built it close for longer than you think you need to.

Then keep them there. Teams tend to relax after a clean first Monday, and the first month end is often when the real surprises arrive: interest runs, statements, regulatory returns and fee cycles that happen only once in thirty days. We plan for hypercare to last through at least one full month end and one quarter end, and we resist every suggestion that it could be shorter.

What we would tell a bank starting today

Start the data work before anything else, because it is always worse than you think, and it is the part you can begin without any vendor at all. Decide early what you will not migrate. Protect the rehearsal schedule from everything, since it is the first thing a late project is tempted to trim and the last thing it can afford to lose.

And be kind to the people doing it. A migration asks a small number of people to carry a great deal of weight for a long time, largely unseen, for an outcome whose whole point is that nobody notices. On that quiet Monday, somebody should notice.

If you are planning a move like this, we would be glad to compare notes. You can read about how we run platforms for banks on the Managed Services page, or talk to the team.

Keep reading

More from the blog

A round steel vault door, closed, set into a stone wall

Security2 min read

What a security questionnaire is really asking

Most vendor assessments run to hundreds of questions, but only a handful decide the answer. Here is what regulated institutions are actually checking for, and how to answer it straight.

Read the article