Search for a command to run...
No comments yet. Be the first to comment.
Most AWS cost optimisation guides ignore PCI-DSS scope, regulatory burst capacity, and payment-specific architecture constraints. Here's what UK fintechs actually need to know.
Most AWS cost optimisation guides ignore PCI-DSS scope, regulatory burst capacity, and payment-specific architecture constraints. Here's what UK fintechs actually need to know.
Most fintechs don't know their end-to-end payment performance until a merchant complains. By then the conversation is already about churn, not optimisation. The Merchant Conversation You Don't Want to
The Diagram That's Delaying Your Audit, Your Acquirer Onboarding, and Your Go-Live
Why Most Fintechs Only Find Out During Due Diligence And How to Get Ahead of It
Built for Payment Engineering Leads
72 posts
Witing about agent-based payment infrastructure — the orchestration patterns, PCI DSS compliance requirements, and failure modes that determine whether your payment system survives production.
Practical guides for engineering teams building on AWS. Tools and assessments at syncyourcloud.io
Why payment state consistency is the architectural problem most engineering teams only take seriously after their first production incident — and what it costs when they do.
Most engineering teams underestimate payment state until it bites them.
Not during the build. During the build, managing payment state feels straightforward. A payment is initiated, processed, confirmed. You store the result. Done.
The complexity surfaces later when your system is under real load, when networks fail at the wrong moment, when a retry fires twice, when a downstream processor times out but doesn't return an error. When a payment is neither clearly successful nor clearly failed, and your system has to decide what to do next.
This is the problem that separates payment infrastructure that scales from payment infrastructure that creates incidents.
A payment isn't a single event. It's a sequence of state transitions, each dependent on the previous, each potentially failing independently.
A typical payment flow:
Initiated → Validated → Authorised → Captured → Settled → Reconciled
In a simple, synchronous system, these transitions happen in sequence, in a single process, with a shared database. If something fails, you roll back. The state is always consistent.
In a distributed system — where validation, authorisation, and settlement involve different services, different databases, and external processors over network calls, consistency is no longer guaranteed. Each transition is a potential failure point. Each failure point is a potential inconsistency.
The question is whether your architecture is designed to handle them correctly when it does.
Your service sends a payment authorisation request to an external processor. The processor receives it, processes it, authorises the payment — and then the network drops before the response reaches you.
From your system's perspective, the request timed out. From the processor's perspective, the payment was authorised.
If your retry logic simply resends the request, you may authorise the payment twice. If you don't retry, you tell the customer the payment failed when it actually succeeded.
Neither outcome is acceptable in a payments context.
What this costs in production: A duplicate authorisation that leads to a duplicate charge triggers a dispute process. At scale — even at 0.1% duplicate rate on 100K monthly transactions — that's 100 disputes per month. At £25-50 dispute handling cost each, that's £2,500-5,000/month in pure overhead before any customer trust damage.
→ Check whether your infrastructure handles lost responses without creating duplicates Run the free Agentic Readiness Assessment →
Your payment processing service successfully captures a payment and needs to update three things: the payment record in your database, the customer's balance, and a downstream ledger service. The first two succeed. The third fails.
Your database says the payment is captured. Your ledger disagrees.
Reconciliation will catch it eventually — but in the meantime, your system is in an inconsistent state. Depending on how your application reads that state, customers may see incorrect balances or receive incorrect notifications.
What this costs in production: Partial write failures that aren't caught by automated reconciliation become manual reconciliation tasks. One engineer spending two days per month untangling ledger inconsistencies is £2,000-3,000/month in engineering cost — before the regulatory risk of inaccurate financial records.
→ Find out what payment state inconsistencies are costing your infrastructure Run the free Payment Risk Estimator →
A payment is processing. Due to a deployment, a crash, or a timeout, the service handling it restarts mid-transition. The payment was in the middle of moving from authorised to captured. When the service restarts, it has no memory of where it was.
Does it retry the capture? Does it check the processor first? Does it assume failure and reverse the authorisation?
The correct answer depends entirely on whether your architecture has explicit state management — or whether it's implicitly relying on everything going right.
What this costs in production: A phantom transition in an agent-based payment system is worse than in a human-initiated flow. An agent will retry automatically, at machine speed, with no human judgment between attempts. Without explicit state management, a single service restart can produce a cascade of phantom transitions across every in-flight payment the agent was handling.
→ Check whether your orchestration produces recoverable failures or unknown states Run the free Agentic Readiness Assessment →
These aren't exotic edge cases. They are normal operating conditions for any payment system at scale. The architecture needs to treat them that way from the start.
Every state transition that crosses a service boundary — including calls to external processors — needs to be idempotent. This means generating a unique idempotency key for each operation and using it consistently across retries. If the same operation is submitted twice with the same key, the system returns the same result without processing it twice.
This is the primary defence against the lost response problem. If you can't tell whether a request succeeded, you retry it with the same idempotency key. The processor handles the deduplication.
Table: idempotency_keys
Partition key: idempotency_key
Attributes: transaction_id, result, status, created_at
TTL: 24 hours
Conditional write: attribute_not_exists(idempotency_key)
Payment state should be modelled explicitly, not inferred. Every valid state a payment can be in, every valid transition between states, and every invalid transition should be defined in code — not scattered across conditional logic throughout the application.
An explicit state machine makes it impossible for a payment to enter an undefined state. It makes the handling of partial failures predictable: you always know what state the payment was in before the failure, and you always know what the valid next steps are.
States: INITIATED → VALIDATED → AUTHORISED → CAPTURED → SETTLED → RECONCILED
Invalid transitions: SETTLED → AUTHORISED, RECONCILED → CAPTURED
Compensation flows: AUTHORISED → VOID (on capture failure)
When a state transition needs to update your database and notify another service, the two operations should not be independent. If your database write succeeds and your service notification fails, you have an inconsistency.
The transactional outbox pattern solves this by writing both the state update and the outbound event to the same database transaction. A separate process reads the outbox and delivers the event reliably. The database transaction either succeeds completely or fails completely — the downstream notification is guaranteed to follow.
Even with all of the above in place, discrepancies will occur. External processors have their own failure modes. Network partitions happen.
Reconciliation — the process of comparing your internal state against your processor's state and resolving differences — is not an afterthought. It is a core part of payment infrastructure.
Reconciliation should run automatically, on a defined schedule, with clear alerting when discrepancies exceed acceptable thresholds. The teams that get this right treat reconciliation as a reliability feature, not a finance operation.
The most common mistake is building payment state management reactively adding idempotency keys after the first duplicate charge incident, adding reconciliation after the first audit finding, adding explicit state machines after the first impossible-state bug.
Each of these is the right fix. But applied reactively, they're applied under pressure, in production, with real customer impact already occurring.
The second most common mistake is underestimating the operational complexity of distributed payment state when making early architectural decisions. Teams that split payment logic across multiple services early before they have the observability, the operational maturity, and the explicit state management to support it often find themselves debugging state inconsistencies that are genuinely difficult to reproduce and fix.
The architecture decisions made at the start of building a payment system determine how hard these problems are to solve later. Getting them right early is significantly cheaper than fixing them under load.
The three failure modes above lost responses, partial writes, phantom transitions are not edge cases. They are guaranteed to occur at scale. The question is whether your infrastructure is designed to handle them or whether you'll discover the gaps in production.
Start with what it's costing you — free, 60 seconds:
The Payment Risk Estimator calculates your monthly infrastructure risk exposure based on your payment service count. It covers duplicate settlement risk, compliance gaps, and manual reconciliation overhead.
Calculate Your Payment Infrastructure Risk →
Then get the full picture — free, 15 minutes:
The Agentic Readiness Assessment covers payment state management across 21 questions — orchestration patterns, idempotency implementation, observability for audit, and failure handling. Scored gap analysis with specific fixes.
Run the free Agentic Readiness Assessment →
If you're building this and need expert support:
Sync Your Cloud membership gives engineering teams access to 26 purpose-built tools for AWS payment infrastructure — including the Agent Flow Simulator for testing payment state transitions without execution risk, the Failure Playbook Generator for documenting recovery procedures, and the full PCI DSS v4.0.1 Gap Analysis.
Plans from £999/month. No execution risk. Full decision logs and evidence packs.
Explore Sync Your Cloud membership →
The full implementation guide covers the exact AWS service stack — Step Functions for state machine orchestration, DynamoDB Streams for the transactional outbox pattern, EventBridge for reliable service notification — that eliminates lost responses, partial writes, and phantom transitions in production.
Read the full implementation guide: Payment State in Distributed Systems →
Sync Your Cloud is the infrastructure readiness platform for engineering teams deploying agent-based payment systems. Built on AWS. Validated for payment infrastructure.