performance
Finding the actual bottleneck rather than the one you assumed.
-
fsGroupChangePolicy: OnRootMismatch and the setgid Invariant That Makes It Sound
Pod stuck in ContainerCreating from fsGroup recursive chown: fsGroupChangePolicy: OnRootMismatch, and why the setgid bit on the volume root makes it safe.
-
Merging Concurrent Polls Without Ever Serving a Half-Built Graph
atomic.Pointer swap of a fully merged snapshot so concurrent polls never serve a torn read, a bug go test -race cannot see. errgroup with timeouts.
-
fsGroup Recursive chown Hangs on Large Volumes
fsGroupChangePolicy: OnRootMismatch stops kubelet re-chowning every file on each mount. Why fsGroup pods restart slowly, and when fsGroupPolicy None applies.
-
Rootless BuildKit: What fuse-overlayfs Costs on a Cold Cache
Rootless BuildKit with --oci-worker-snapshotter=fuse-overlayfs needs no --privileged, but cold-cache builds pay a FUSE round-trip per file. Measuring it.
-
Running Ceph on Commodity Hardware: Failure Domains and the Four Things That Bite
ceph osd df --format json: alert on the fullest OSD, not the mean, because one full OSD stops writes. Plus CRUSH failure domains and monitor quorum.
-
Serving Several Local Models on One GPU with On-Demand Loading
llama-swap config.yaml with cmd, ttl and healthCheckTimeout: loading llama-server models on demand and unloading idle ones on a single GPU.
-
Sizing One GPU for Local LLM Inference: VRAM, Quantisation and KV Cache
llama-bench -ctk q8_0 -ctv q8_0 vs f16: estimating VRAM for GGUF weights plus a KV cache that grows with context, and catching slow KV cache types.
-
Preventing Overselling with a Database CHECK Constraint
violates check constraint "stock_not_negative": why check-then-decrement oversells under READ COMMITTED, and a single conditional UPDATE that cannot.
-
A Shared Multi-Stage Base Image for a C++ Service Fleet
vcpkg install pistache rebuilt in every service Dockerfile: a shared, versioned multi-stage base image that compiles C++ dependencies once for the fleet.
-
Migrating Redis Consumers from Python to C++ on Constrained Hardware
redis-plus-plus xreadgroup times out before its block timeout: set socket_timeout to 0. Replacing idle Python Redis stream consumers with C++.
-
Sizing Redis maxmemory Against a Stream Trim Threshold
OOM command not allowed when used memory exceeds maxmemory on XADD MAXLEN: why the stream never trims itself, and why a standalone XTRIM still works.
-
Tuning Ceph Recovery: The Trade Between Degraded Time and Client Latency
osd_max_backfills does nothing under mClock unless osd_mclock_override_recovery_settings is set. Choosing and measuring osd_mclock_profile for recovery.
-
Capacity Planning Against the Fullest OSD, Not the Average
Ceph full_ratio applies per OSD: one OSD at 95% stops writes while ceph df shows 70%. Alerting on the fullest OSD in ceph osd df, plus the upmap balancer.
-
CRUSH Hybrid Rules: Pinning the Read Primary to SSD Without Rebalancing
step take default class ssd, chooseleaf firstn 1, then class hdd firstn 0: a CRUSH rule with an SSD primary and HDD replicas, tested with crushtool.
-
A Layered Storage Benchmark: Isolating Device, Network and Protocol Bottlenecks
fio, iperf3, rados bench, then fio on RBD: benchmarking device, network, Ceph protocol and full stack separately to find which layer caps throughput.
-
Multi-Stage Docker Builds: Shipping a Runtime Without the Toolchain
COPY --from=build with CGO_ENABLED=0 and distroless/static: a multi-stage Dockerfile taking a Go image from 812MB to 24MB, and the no-shell trade-off.