A financial services firm ensures reliable AI data delivery

A leading European asset manager learned the hard way that object storage alone cannot absorb the concurrency of AI workloads. After a client misconfiguration triggered a denial-of-service event, the firm deployed F5 solutions to ensure reliable, scalable, and high-performance S3 data delivery for AI model training.

Challenge

A European asset manager with a strong presence in financial services and real estate is investing heavily in AI. The company’s data science teams train models on large datasets stored in Dell Elastic Cloud Storage (ECS), an S3-compatible object store, with AI workloads running on Kubernetes. The weak point was the path between those workloads and the storage itself.

Object storage platforms are engineered to scale capacity, not to police traffic. For decades, applications connected with storage over dedicated protocols on controlled networks, so no traffic management was needed in front of a filer or an array. S3 storage changed that. Object storage is now accessed over HTTP by any client holding a key, which means the storage front end behaves like a web service and can fail in the same ways. Individual storage nodes have practical limits on how many concurrent TCP connections they can handle. When AI pipelines open thousands of parallel sessions to pull training data, those limits can be reached quickly, and storage teams rarely see the problem coming.

For the financial services firm, a misconfigured client opened more concurrent connections than the origin could handle, effectively causing a denial-of-service condition against the company’s own storage. AI data ingestion stopped.

The incident made two things clear to the infrastructure and storage teams.

First, a single misbehaving client should never be able to take down a shared data platform. Particularly in financial services, reliability is not optional. Model training schedules, analytics, and downstream business decisions all depend on the data platform remaining in operation.

Second, in light of company plans to add ECS nodes to feed growing AI initiatives, the problem would only compound. Moving forward required engineering discipline at the storage boundary, with a control point that could shape, balance, and protect S3 traffic before it ever reached a storage node.

Solutions

The firm evaluated several potential solutions, including HAProxy and options from Kemp Technologies, but found them limited in application delivery features tailored to the S3 protocol.

Early proof-of-value demonstrations became decisive, displaying the ability of the F5 solution to handle S3 protocol challenges such as session persistence while integrating cleanly with ECS. Dell, a technical partner both for the company and F5, played a pivotal role by aligning the F5 solution with the ECS infrastructure and advocating for the joint architecture.

Ultimately, the financial services firm selected F5, deploying four F5 rSeries r5600 appliances running BIG-IP Local Traffic Manager (LTM) as a dedicated delivery tier in front of the ECS storage.

The deployment followed a proven reference design, which reduced risk and provides a roadmap as the AI estate matures. The resulting dual data center, active/active architecture is deliberately simple. It leverages dynamic routing mechanisms to provide horizontal scaling and high availability without DNS dependency. Every S3 request from the Kubernetes-hosted AI workloads now passes through BIG-IP LTM before touching storage. That single design decision resolves the original failure mode and establishes a foundation for growth.

Results

Connection pooling addresses the concurrency problem directly, so the incident that started this project cannot recur. BIG-IP-LTM manages the flood of client-side TCP connections and shapes them into a smaller, controlled set of server-side connections to the ECS nodes. It also enforces connection discipline so a single misconfigured or aggressive client can no longer exhaust a storage node. The ECS platform stays available to every team even when one client misbehaves.

Deliver measured performance gains

Consistent URL hashing keeps request distribution intelligent rather than random. Requests for the same objects are steered predictably across the ECS cluster, which improves cache utilization, avoids hot spots, and keeps throughput stable as storage nodes are added.

In addition, internal performance tests with the Mongoose storage performance testing tool demonstrated a 500% performance gain on processed inputs and outputs after implementing F5 solutions in the data path. By optimizing TCP connections, multiplexing server-side requests, and leveraging dynamic routing, the platform delivered immediate throughput improvements for demanding AI workloads.

Scale AI initiatives without rearchitecting

The firm plans to grow its ECS environment as AI adoption expands. Because BIG-IP LTM abstracts storage endpoints behind a single virtual address, the firm can add nodes to scale the ECS footprint without touching application configurations or disrupting workloads. In this way, the architecture that fixed a reliability problem also enables growth.

Turn storage into an architected service

Placing a dedicated application delivery tier in front of storage supports good engineering practices. Traffic is observable, balanced, and governed in one place instead of being an unmanaged, tightly coupled mesh of clients hitting nodes directly. For the storage architect, this means predictable node behavior and clean capacity planning. For the CIO, it means the AI data platform is a well-engineered system rather than an accumulation of point-to-point connections.

Build on an AI data delivery partnership

F5 and Dell continue to engineer jointly for exactly this kind of workload, pairing BIG-IP LTM traffic management with S3-compatible object storage for AI pipelines. The lesson for other enterprises is that S3 object storage is a web service and AI workloads will stress it like one. A dedicated application delivery controller in front of the storage can turn an exposed origin into the resilient data pipeline required for effective AI implementation.