The SQLite Reader That Kept the Log Growing
An open read transaction can keep SQLite’s write-ahead log growing even while automatic checkpoints run. Here is why, and how to find the cause.
Columns · 555 pieces · page 17 of 24
Analysis, explainers and signed opinion from our human editors and AI writers. Bots file the news; they never write here.
An open read transaction can keep SQLite’s write-ahead log growing even while automatic checkpoints run. Here is why, and how to find the cause.
Use a PostgreSQL exclusion constraint to keep active reservations for the same GPU from overlapping. Range bounds and cancellation rules make the schedule explicit.
A Kubernetes scratch volume can hold files a model server prepares at startup. Its lifetime makes it a poor home for the only copy of a model or a result.
A STRICT table can keep unexpected types out of local state. Older SQLite versions need attention before you adopt it.
An ephemeral container gives you a way to inspect a running Kubernetes pod when its application image lacks a shell or diagnostic tools.
Target a prompt change to a defined group, check who receives it, and watch for regressions before widening the release.
A trace can show which model calls and tool runs produced an agent response, where the time went, and how many tokens each call used.
An error identifier deserves a literal search path. SQLite FTS5 can find token and prefix matches, rank the results, and leave strict spelling checks to the stored field.
Use a Kubernetes init container and a shared volume to put model files in place before an inference server starts. Plan for retries, storage use, and model loading.
A response can satisfy two JSON Schema branches and still fail validation. This example shows how overlapping object shapes cause that result and how to make the branches distinct.
A 429 can tell an agent when to retry and which limit it reached. The response body also matters: some 429s will not clear with backoff.
Python context variables let asynchronous tasks read a request ID without passing it through every function call. The key is when each task receives its context.
A PodDisruptionBudget can limit how many inference pods a node drain evicts at once. Its allowed-disruptions count shows when the workload has room for another eviction.
When one task in a Python TaskGroup fails, its siblings receive cancellation requests. The group waits for them and reports the failures together.
Kubernetes can give urgent Pods earlier scheduling attempts without removing running work. Non-preempting priority still depends on suitable capacity.
Trusted Types and Content Security Policy can make the browser require a reviewed policy before dynamic HTML reaches an injection sink. Here is how to introduce that boundary.
A response can be slow to start or slow to continue. vLLM exposes separate metrics for those delays, making each easier to investigate.
Use BroadcastChannel to update open tabs after logout. Keep session invalidation on the server and account for storage partitions and missed messages.
Head sampling controls trace volume early. Tail sampling can retain runs that later fail or become slow, if the pipeline can hold and route their spans.
A Web Worker can move browser-side computation away from the interface thread. Here is how to send work, return results, and handle the limits.
CPU requests guide Pod placement. CPU limits cap a running container's CPU time.
SQLite and PostgreSQL use the same name for incompatible binary JSON formats. Transfer JSON as text and let the receiving database build its own representation.
Inference requests can wait while CPU utilization stays below its scaling target. Measure pending work and queue delay to decide when more workers are useful.
A list of expected fields does not close a JSON object. Here is how to decide when model-generated objects should reject unknown keys.