0

Cross-Tenant Throttling Incidents

4.7x

Peak Concurrent Workflow Capacity

Portrait of Tobias Reyes, Director of Platform Engineering at Loomwork

Tobias Reyes,

Tobias Reyes

Director of Platform Engineering

Isolated architectural cylinders representing strict tenant-scoped quota management to prevent noisy-neighbor API throttling.

The Problem: Shared Rate Limits Created Cross-Tenant Contention

Loomwork’s automation platform executes customer-defined workflows that frequently include LLM-powered steps, such as classification or summarization actions inserted into a Zap-style chain. All outbound LLM calls originally shared a single rate-limited API key per model provider, which meant a single high-volume tenant running a bulk workflow could exhaust the shared quota and cause HTTP 429 responses for unrelated tenants’ workflows.

The architectural flaw was the absence of per-tenant request budgets at the infrastructure layer; rate limiting existed only at the provider level, which has no concept of Loomwork’s internal tenant boundaries. Support tickets attributing failed workflow runs to ‘random’ API errors averaged 60 per month.

The Implementation: Meridian Tenant-Scoped Quota Management

Loomwork integrated Meridian’s tenant-scoped quota layer, which sits between Loomwork’s workflow executor and upstream model providers. Each tenant is assigned an independent token-bucket budget, configurable per plan tier, and enforced at the edge before a request ever reaches the shared upstream connection pool.

We needed rate limiting that understood our data model, not just requests-per-second against a single API key. Meridian let us express ‘tenant’ as a first-class routing dimension.

When a tenant’s bucket is exhausted, Meridian queues the request with a bounded wait rather than failing immediately, and surfaces a distinct error code so Loomwork’s executor can differentiate tenant-quota exhaustion from genuine upstream provider failure, eliminating the ambiguity that previously generated misleading support tickets.

Results: Measured Infrastructure Impact

Cross-tenant throttling incidents dropped to zero in the six months following deployment. Because tenant budgets are now isolated and independently scalable, Loomwork increased peak concurrent workflow execution capacity by 4.7x without any change to upstream provider contracts, and support tickets related to unexplained 429 errors fell to near zero.

Ready to route your first payload?

Get your first API key and start routing production traffic today.

Aquire $129

Aquire $129

Aquire $129

Create a free website with Framer, the website builder loved by startups, designers and agencies.