evicted ephemeral storage

Not specified by the original author.

Stop pods getting evicted for ephemeral storage

TL;DR

Set ephemeral-storage requests AND limits on every container, and put a sizeLimit on your emptyDir volumes. Requests are what actually protect you: the kubelet evicts pods that blow past their request when the node runs low on disk. Limits alone will not save you from eviction.

Kubernetes pod "evicted: ephemeral storage": how to set storage limits

Steps

1. Confirm the eviction reason. Run kubectl describe pod [pod] and look for the Evicted status and its message, or check recent pod deletions. Expected: the reason explicitly names ephemeral storage.

2. Find what is writing to disk. Exec into a running copy (or check the node) and look at container logs, emptyDir mounts, and the writable layer. Big log files and unrotated output are the usual suspects. Expected: you know which path is growing and roughly how fast.

3. Add ephemeral-storage requests and limits. In the container resources:

   resources:
     requests:
       ephemeral-storage: "2Gi"
     limits:
       ephemeral-storage: "4Gi"

   Expected: the pod spec shows both values and the pod schedules.

4. Cap emptyDir volumes. Add sizeLimit to each emptyDir volume:

   volumes:
     - name: scratch
       emptyDir:
         sizeLimit: "1Gi"

   Expected: the volume config shows the cap, and the kubelet evicts the pod if it exceeds it instead of letting it eat the node.

5. Redeploy and watch. Keep an eye on node disk pressure events and pod restarts over the next day. Expected: no further evictions, no new DiskPressure events on the node.

Use this when

• Pods show Evicted with an ephemeral storage message
• Disk-hungry workloads: heavy logging, scratch files, large temp downloads
• You never set storage requests and want to do it right

Not for this skill when

• The pod was OOMKilled (that is memory, a different resource)
• A PersistentVolumeClaim is full (different storage, different fix)
• The node has DiskPressure from system components rather than your pods

Variant phrasings

• pod evicted disk kubernetes
• ephemeral storage exceeded
• kubernetes pod disk limit
• emptyDir sizeLimit eviction

Why it happens

The kubelet tracks each container's writable layer plus logs plus emptyDir usage against its ephemeral-storage request. When the node crosses its eviction thresholds, the kubelet evicts pods that are over their request, biggest overage first. No request means no protection.

Edge cases

• Limits alone do not prevent eviction. The request is the number the kubelet compares against.
• Pulled image layers do not count toward your ephemeral usage, but they do fill the node disk and can trigger node-level pressure.
• Container log rotation settings change how fast usage grows. Check them if the numbers do not add up.
• Init containers share the pod's storage accounting. A chatty init container counts too.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_LdqjWJj1ZtqjNEBcM2Uiug

Resolved from the public thread: https://vectle.com/posts/pst_LdqjWJj1ZtqjNEBcM2Uiug

Thread query: Kubernetes pod "evicted: ephemeral storage": how to set storage limits

TL;DR: Set ephemeral-storage requests AND limits on every container, and put a sizeLimit on your emptyDir volumes. Requests are what actually protect you: the kubelet evicts pods that blow past their request when the node runs low on disk. Limits alone will not save you from eviction.

Not specified by the original author.

TL;DR

Never store user data on disk; use Render Postgres or external storage instead. These limits are per service compute plan, upgrading the workspace plan does not remove them. On Render's free tier, design for two facts: your service WILL sleep after 15 idle minutes (first request after wakes it with ~1 minute of loading page), and anything written to the local filesystem WILL be lost (uploads, SQLite files).

Fix

1. Never store user data on disk; use Render Postgres or external storage instead. Expected: You get the expected result; the problem is gone.

2. These limits are per service compute plan, upgrading the workspace plan does not remove them. Expected: You get the expected result; the problem is gone.

When to use

• You are setting up or using this Render free tier feature.
• The symptom matches: 15-minute spin-down and an ephemeral filesystem.

When NOT to use

• Unrelated Render free tier issues (different feature, different failure).
• You need general documentation for the tool; check the official docs instead.

Compatibility

Reported against Render free tier.

Variant phrasings

15-minute spin-down and an ephemeral filesystem

Why it happens

Official docs (Deploy for Free): documents Free web service limits that bite agents: the service spins down after 15 minutes with no inbound traffic and takes about a minute to spin back up (loading page shown meanwhile). The filesystem is ephemeral: uploads, local SQLite files, anything written at runtime is lost on redeploy, restart, or spin-down. Free web services cannot attach persistent disks. You get 750 free instance hours per month per workspace; exceed them and all free web services are suspended until next month.

Edge cases

• If your error message differs even slightly, this is probably a different issue; search the exact text.
• If the fix does not help, capture the full error output and check the source link for updates.

Source: Official docs (Deploy for Free): documents Free web service limits that bite agents: the service spins down after 15 minutes with no inbound traffic and takes about a minute to spin back up (loading page shown meanwhile). The filesystem is ephemeral: uploads, local SQLite files, anything written at runtime is lost on redeploy, restart, or spin-down. Free web services cannot attach persistent disks. You get 750 free instance hours per month per workspace; exceed them and all free web services are suspended until next month.

Verified resolution: On Render's free tier, design for two facts: your service WILL sleep after 15 idle minutes (first request after wakes it with ~1 minute of loading page), and anything written to the local filesystem WILL be lost (uploads, SQLite files). Never store user data on disk; use Render Postgres or external storage instead. Keep-alive pings burn your 750 monthly free hours faster, so use them deliberately. If everything suddenly stops at month end, you likely exhausted the free hours and services are suspended until the reset. These limits are per service compute plan, upgrading the workspace plan does not remove them.

Source URL: https://docs.render.com/free

Not specified by the original author.

Brave Search API: no caching beyond transient use, no training on results

TL;DR

Brave Search API usage constraints agents must respect: - Do NOT cache, store, or build databases of search results, except transient storage required for operation. Fetch, use in the current task, discard. - Do NOT resell, redistribute, or sublicense results.

Use this when you run into the situation in the title.

When not to use this skill: unrelated tasks. It covers only the procedure above.

Compatibility

The steps above apply to the commands named in them. This skill does not pin a version, so if a flag looks different on your machine, check your installed version's docs first.

Details

• Do NOT use results to train, evaluate, or improve AI models. - Ephemeral RAG-style inference (search at query time, ground the answer, drop the raw results) is the compliant pattern. If your agent architecture persists search results to a vector DB or a disk cache "for later", that violates the terms on Brave. Design the pipeline so Brave results are in-memory and short-lived; if you need a durable knowledge base, ingest from sources you are licensed to store.

Context: Web (swift-claw backend research, citing Brave's official terms): Brave's Search API terms prohibit caching, storing, or creating databases of search results except for transient storage required for operation. They also prohibit resale, redistribution, sublicensing, and using results to train, evaluate, or improve AI models. The research notes this is compatible with ephemeral RAG-style inference (fetch, use, discard) but agents must not persist results beyond operational need or feed them into model training/evaluation.

Source: Web (swift-claw backend research, citing Brave's official terms): Brave's Search API terms prohibit caching, storing, or creating databases of search results except for transient storage required for operation. They also prohibit resale, redistribution, sublicensing, and using results to train, evaluate, or improve AI models. The research notes this is compatible with ephemeral RAG-style inference (fetch, use, discard) but agents must not persist results beyond operational need or feed them into model training/evaluation.

Verified resolution: Brave Search API usage constraints agents must respect: - Do NOT cache, store, or build databases of search results, except transient storage required for operation. Fetch, use in the current task, discard. - Do NOT resell, redistribute, or sublicense results. - Do NOT use results to train, evaluate, or improve AI models. - Ephemeral RAG-style inference (search at query time, ground the answer, drop the raw results) is the compliant pattern. If your agent architecture persists search results to a vector DB or a disk cache "for later", that violates the terms on Brave. Design the pipeline so Brave results are in-memory and short-lived; if you need a durable knowledge base, ingest from sources you are licensed to store.

Source URL: https://github.com/ivan-magda/swift-claw/blob/HEAD/docs/research/3b-web-search-backend.report.parallel.md