Skip to content
Notifications
Clear all

Deployed LangSmith to 200 users across 4 time zones - stability issues

2 Posts
2 Users
0 Reactions
78 Views
(@fionac)
Reputable Member
Joined: 4 months ago
Posts: 186
Topic starter   [#16800]

We completed a company-wide rollout of LangSmith last week to our internal operations teams. The goal was to give our support, sales, and onboarding specialists a unified platform to build and refine their LLM prompts for customer interactions. In total, we have about 200 users now accessing it across four primary time zones (EST, CET, IST, SGT).

The deployment itself went smoothly, but we're hitting some unexpected stability walls under real load. The main issue seems to be with the web interface becoming very slow or unresponsive during peak hours in each region, particularly around 10 AM local time when teams are starting their day and pulling up their workspaces. We've had multiple reports of timeouts when trying to view traces or dataset runs. It's not a complete outage, but the latency is disrupting workflows.

From a configuration standpoint, we're using the cloud-hosted version, not self-managed. Our setup involves each team having its own project, with shared datasets for benchmarking. We haven't done anything too complex with webhooks or evaluations at scale yet, so I'm surprised by the performance hiccups.

I'm curious if others have managed similar-scale deployments. Did you run into similar growing pains with the UI performance? We're looking into whether we need to stagger some usage or if there are specific settings we might have overlooked. Our initial focus was on access controls and project structure, assuming the infrastructure would handle the concurrency.



   
Quote
(@code_reviewer_anna_v2)
Honorable Member
Joined: 7 months ago
Posts: 422
 

I had a similar experience when we rolled out LangSmith to about 150 users. The web interface latency during regional peaks was the biggest headache. It turned out the default project views loading *all* recent traces were the culprit.

Are you monitoring your database connection pool? At that scale, the cloud instance might be hitting connection limits when everyone logs in at 10 AM. We added some aggressive client-side caching for trace lists, which helped a lot. Maybe start by checking if the slowdowns correlate with specific endpoints, like the `/traces` or `/runs` APIs.

Did you consider using read replicas for the trace data? Even on the cloud plan, sometimes you need to nudge support about scaling the underlying DB.


Clean code, happy life


   
ReplyQuote