< Insights

Latency, scale, and complex integrations: the invisible bottleneck of orders and allocations

  • Cloud & Infrastructure
  • article

In trading desks and investment platforms, the discussion about latency often stays within the technology area and is treated merely as an infrastructure metric. However, this is a framing error; when the delay starts to affect the executed price, asset allocation, and risk assessment, it ceases to be a technical indicator and becomes a line item on the income statement. Below are five points that define where this line is crossed and what needs to change in the architecture to support real-time decision-making.

When does latency stop being a technical problem and become an execution loss?

The boundary is easy to identify: latency becomes a loss when the system’s response time is greater than the time during which the information remains valid. Beyond this limit, the system is not only slow; it is making decisions based on data that no longer describes the market.

The effect does not appear as an incident. It manifests as recurring slippage, partially executed orders at worse prices, and a systematic difference between the decision’s reference price and the average price obtained. These are small losses per operation but significant when accumulated over the year, and they are almost never attributed to data architecture.

The practical test is to measure the interval between the market event and the decision made about it, and compare that number with the window during which the opportunity is valid. If the first is greater than the second, there is an ongoing loss that no one is accounting for.

Why do point-to-point integrations stop scaling before the volume?

The point-to-point model works well initially but fails due to a structural reason. With each new system connected, the number of integrations grows at a faster rate than the number of systems. Ten fully connected systems require dozens of integrations, each with its own contract, error handling, and maintenance cycle.

The practical impact is related to time, not throughput. Adding a new broker, a new asset class, or a new custodian stops being a configuration task and turns into a project. The technology area becomes a bottleneck for business decisions, and the institution loses speed in entering markets where the window is short.

There is also a silent effect: each integration carries its own version of the same information. When two routes diverge, no one knows which is correct without manual investigation. The cost appears in the back office, but originates in the architecture.

What does intelligent order routing require from the architecture?

Intelligent routing means choosing with updated information, within a time limit, without forgoing the record of what was decided.

This imposes three requirements. The first is access to market and position data with latency compatible with the decision, which means consuming by event and not querying a transactional database. The second is a clear separation between decision logic and connection with each destination, so that adding or removing a route does not require altering the engine. The third is complete traceability: what information was available, what alternatives were considered, and why that route was chosen.

The third requirement is often treated as a compliance requirement and is, in fact, a condition for improving the routing itself. Without a record, there is no way to evaluate the quality of decisions or calibrate the criteria over time.

Real-time allocation and risk: why both need the same data

Allocation and risk are treated as distinct processes, with separate systems, and this separation creates an inconsistency that is difficult to resolve afterward. Allocation distributes execution among portfolios. Risk evaluates exposure, limits, and framing. Both rely on the same information: what the position is now.

When each one works with its own database, the institution operates with two simultaneous truths. The allocation engine distributes an order that the risk system has not yet seen. The limit control approves an exposure calculated based on the closing position. Neither is wrong in isolation, but together, they produce inconsistent decisions.

The correction is not to integrate the two systems with one more route. It is to establish a single source of position, updated by event, that both consult. From there, allocation and risk may disagree based on criteria, not on data lag, and this is a divergence that can be resolved by rules, not by investigation.

What changes when the decision occurs at the event, not at the closing?

Processing by event changes the unit of work of the operation. Instead of accumulating records and deciding on the set, the system decides on each fact at the moment it occurs.

Three consequences matter for the business. Control becomes preventive: the limit is verified before execution, not identified later in reconciliation. The error no longer contaminates the batch because the failure remains isolated in the event and can be reprocessed alone. And the audit trail becomes a byproduct of the operation, not a report constructed afterward.

The trade-off is that the model requires a technical discipline that the batch model did not, mainly idempotency and ordering control. Without this, real-time generates inconsistency faster than the batch generated delays.

Success case 

A large credit union operated about 50 critical processes on a legacy automation platform. The bottleneck was not volume: each new process became expensive and slow, and governance did not meet the expectations of a regulated financial operation.

With Objective, the credit union built a cloud orchestration layer that began coordinating the existing automation agents instead of replacing them. In four months, the first processes went into production. The full article details how the architecture was designed and which processes migrated first. Read the success case.

None of these changes require replacing the core. They require choosing a flow where latency is already costing money, measuring the interval between the event and the decision made about it, and treating that number as a business indicator, not just an infrastructure one. From there, the discussion about architecture shifts from technical to how much execution the operation is losing each month.

Objective operates in mission-critical environments in the financial sector, combining event-driven architecture, data engineering, and legacy system modernization, without interrupting operations. Talk to our specialists.

Frequently asked questions on the topic

What latency is acceptable?

There is no universal number. The parameter is the window during which the information remains valid for that decision. Above that, there is loss; below it, investing in reduction tends not to pay off.

How many integrations justify switching from the point-to-point model?

The signal is not the quantity; it is the time. When adding a new broker or asset class stops being configuration and becomes a project, the model has already crossed the line.

Does an event bus replace the core?

No. The core continues to function as the recording system. The bus is where the information circulates and the decision happens.

Does intelligent routing require AI?

Not necessarily. It requires updated data, separation between logic and connection, and recording of the decision. AI comes in later to calibrate criteria based on that history.

What is the first step?

Measure the interval between the market event and the decision made about it in a specific flow, and compare it with the validity window of that information.

Insights from our team

Get insights from our team of experts on software development methodologies, languages, technology and more to support your team in business operation and strategy.
Learn more about OGI
Ask us about our expertise, services, and consulting.