Cases13 cases

8 to 12 of 13

1. IL-0024 Expensive cache-backed API calls and uncancelled database requests overloaded MongoDB and took the Hugging Face Hub down three times

   Hugging Face2024

   Symptoms

   At 08:45 CET on 22 April 2024 the Hugging Face Hub became unstable, and by 09:30 it was unresponsive. For most users the website was slow or could not be reached at all.

   Service outageHigh latencyDatabase overload

   Trigger

   Bursts of requests to revision endpoints that read heavily from the *_info_cache collections: a call to https://huggingface.co/api/datasets/lighteval/mmlu/revision/main rose sharply right before the first crash, and a call to https://huggingface.co/api/models/sentence-transformers/all-mpnet-base-v2/revision/main, a model with 300+ associated spaces, lined up with the crash at 21:30 on 23 April.

   Root cause

   Software defect

   Component

   Application databaseModel hubCluster autoscaler

   Contributing factor

   No headroomRetry stormOverlapping failures

   Detection

   Instability started at 08:45 on 22 April and the Hub was unresponsive by 09:30; the team followed MongoDB Atlas metrics and a MongoDB support case. The root cause was found days later from a new graph of requests per URL.

   Remediation

   • Added a firewall rule blocking all requests with a Java/* user agent
   • Disabled the quicksearch feature on the Hub
   • Scaled the MongoDB Atlas cluster by hand from tier 2 to tier 3
   • Switched the Hub to maintenance mode with a 503 page through AWS WAF and blocked incoming traffic
   • Scaled all Hub pods down to 0 and brought them back
   • Set an aggressive rate limit
   • Added timeouts to requests and reworked database queries that were slow or returned large bodies

   Action items

   • Add timeouts to requests based on their usual completion time
   • Optimise requests that took too long or returned a large body, and rework poorly designed database queries
   • Remove old unused indexes from the database
   • Build new dashboards and monitoring metrics
   • Move the *_info_cache collections to Redis
   • Make it possible to abort any database request when the client aborts theirs

   Lessons

   Finding the root cause took two to three days: the existing charts did not show it, and a per-URL request breakdown did. Not cancelling database work when a client goes away turns a strained database into a thundering herd. A close inspection of the database exposed architectural flaws and bottlenecks, such as a collection with large unused indexes, and taught the team which metrics need closer watching to anticipate crashes. Unrelated failures happening at the same time added to the stress and made restoring service harder.

   View case
2. IL-0123 A response buffer allocated anew in a loop made OpenAI's routing layer nodes hit their memory limits and fail readiness checks on a morning with more completions than any day before, causing 502 and 503 errors across all models

   OpenAI2023

   Symptoms

   Between 5:42 AM and 7:16 AM PT on 8 November 2023 a large share of requests to OpenAI failed with 502 or 503 errors. Every model and API endpoint saw significant failures.

   High error rateService outageProbe failures

   Trigger

   A morning with dramatically more completions than any previous day pushed the routing layer nodes, which already had a chronic memory problem, past their memory limits.

   Root cause

   Software defect

   Component

   Model request routingContainer memory limit

   Contributing factor

   No alert on limitNo headroom

   Remediation

   • limited incoming traffic
   • redeployed the routing service en masse
   • ramped traffic back up slowly

   Action items

   • Pre-allocate the responseBuffer and reuse it instead of allocating a new one in a loop (3X improvement in memory and CPU usage)
   • Adjust the configured memory limits to an appropriate level so the service has significant headroom
   • Add rate limit controls to shed traffic more gracefully
   • Increase the service's capacity
   • Change alerting to catch the underlying memory behaviour before it escalates into service issues
   • Configure autoscaling for the service, which was not possible before because added capacity could hurt an upstream service

   View case
3. IL-0070 Critical pods on core nodes evicted when Prometheus began scraping about 200 nodes after a dask-gateway scale-up

   2i2c2023

   Symptoms

   The hub pod of the prod JupyterHub on the leap cluster (leap.2i2c.org) restarted unexpectedly during a large node scale-up, and a user opened a support ticket. The ingress-nginx and proxy pods kept running.

   Pod evictionPods Pending

   Trigger

   A user's dask-gateway workload added about 100 dask-worker nodes within a minute, and Prometheus started scraping the metrics exporters on about 200 nodes.

   Root cause

   Capacity exhaustion

   Component

   PrometheusNode pool

   Contributing factor

   No headroomExcessive blast radiusMonitoring blind spot

   Detection

   A user support ticket (2i2c Freshdesk ticket 414) about the leap hub; the engineer then looked at Grafana and the cluster's nodes and events.

   Action items

   • Configure CPU and memory requests for the hub and proxy pods (#2127; at least 100m CPU was proposed)
   • Make the Grafana OOMKiller dashboard work (#2213)
   • Move LEAP cloud infrastructure to shared nodes (#2209)
   • Update resource requests and limits, mainly for ingress-nginx and Prometheus (#2324)
   • Decide whether Prometheus should scrape dask worker pods (#2366)

   Lessons

   Prometheus load grows with the number of nodes, so a burst of autoscaled nodes can push a nearly full core node over the edge and evict the pods that matter most. Core pods that request 10m of CPU get a tenth of the CPU share of busy neighbours that request 100m. Kubernetes events expire after an hour, so evidence of OOM kills needs a dashboard that keeps it.

   View case
4. IL-0115 One of OpenAI's three production AKS clusters went offline through a human error, and rebuilding it took 10½ hours because an AKS Kubernetes upgrade broke the ingress and AKS cannot be downgraded

   OpenAI2022

   Symptoms

   At 18:56 PDT on 23 June 2022 (01:56 UTC on 24 June) several models went offline at once: some fine-tuned curie and davinci models, the Codex beta models code-davinci-001 and code-cushman-001, the legacy GPT models curie, babbage and ada, and all Embeddings models.

   Service outageHigh error rateIntermittent timeouts

   Trigger

   A destructive operation on one of the three production clusters, caused by human error that the existing safeguards did not catch, took the whole cluster offline.

   Root cause

   Operator error

   Component

   Ingress controllerModel hub

   Contributing factor

   No safeguard on destructive changeNo headroomDeferred upgradesOverlapping failures

   Remediation

   • re-allocated the models from the lost cluster to the remaining two clusters
   • rebuilt the VPN link with Azure support after losing the connection to AKS during cluster setup
   • built workarounds with Azure support after losing the connection to Azure Private Link
   • tore down the cluster and rebuilt it from scratch after the upgraded Kubernetes version broke the ingress
   • moved traffic back into the rebuilt cluster between 05:30 and 06:30 PDT

   Action items

   • Add checks and locks so that this and similar production-critical infrastructure cannot be modified by accident
   • Add better visibility into the operations being performed and make parts of the system more lexicographically distinct
   • Keep working with Azure on the AKS network connectivity issues so they do not disrupt new cluster operations
   • Update playbooks with the new cluster failure modes and change the setup steps for new clusters and infrastructure
   • Upgrade Kubernetes systematically and set aside time so the clusters do not fall behind again
   • Shard fine-tuned models across multiple clusters and work towards more automatic cross-cluster replication
   • Distribute and replicate Azure Blob Storage so fine-tuned models load faster and more reliably
   • Finish rolling out node-to-node caching to relieve Azure Blob Storage bottlenecks
   • Change how model weights are stored so they are more efficient and faster to load

   Lessons

   Losing one of three clusters exhausted the GPU headroom, and GPUs could not be added on short notice, so full recovery depended on bringing the same cluster back. Rebuilding it exposed a deferred Kubernetes upgrade: the newer version broke the ingress, and because AKS cannot downgrade there was no way back except a full rebuild. Recovering many large models at once is limited by the bandwidth of the storage they load from, and the caches that normally hide this do not help when everything moves to a different cluster.

   View case
5. IL-0126 The OPT-175B training run restarted more than 40 times in two weeks as cloud GPU nodes failed with ECC errors, replacement nodes came back faulty, and checkpoint downloads hung

   Meta AI2021

   Symptoms

   Between the team's 17 November 2021 update and 3 December 2021, the 175B-parameter OPT training run went down and had to be restarted more than 40 times, for hardware, infrastructure and training stability reasons. A hardware fault could stop the run for hours at any time of day, and several did over the Thanksgiving break.

   Training job failureGPU unavailable

   Trigger

   Nodes in the cloud training cluster failed with ECC errors, and each failure took the training run down.

   Root cause

   Dependency failure

   Component

   GPU hardware and memoryNCCLInfiniBand fabricObject storage

   Contributing factor

   No headroomLimited debug accessNo runbook

   Detection

   During these two weeks the team added monitoring of train.log, GPU burn-in testing to detect memory errors, and automated InfiniBand testing.

   Remediation

   • replaced failed nodes, and later did it through a script
   • added GPU burn-in testing to detect memory errors
   • automated InfiniBand testing
   • added monitoring of train.log
   • debugged the cloud provider's Python API until checkpoint downloads from the blob store worked well enough

   Action items

   • Run an on-call rotation for the training run, with runbooks and tooling for the most common issues, and invite more people to join it

   Lessons

   At this scale cloud hardware failures are certain, and the team concluded that automating the handling of hardware failures is necessary. Without spare nodes and burn-in tests, each failure costs hours and a replacement can bring the same fault back. The team also concluded that cloud infrastructure instability is something to always prepare for, because the customer has no control over the underlying hardware. A checkpoint large enough to need hundreds of multi-gigabyte downloads makes the storage path part of recovery time.

   View case

By root cause8