The fd files under /proc/<pid>/fd/ are created when a user reads or lists the fd files, not when the actual files are opened. The inode of a fd file is not deleted when the file is closed. This inode is disconnected with the closed file and stays in the VFS dcache.

Some tips and findings on tracing go binaries using bpftrace.

I went through glibc malloc code when troubleshooting a segfault. Here are some notes on the malloc code.

Requesting CPU resource in a pod spec does not guarantee the requested CPU for a container. The only way to guarantee this is to use "static CPU manager policy" and exclusively allocate CPUs to the container.

I dived into the kernel scheduler code under kernel/sched/ to understand how CFS works and how the task group cpu.shares value is used in CFS.

Many changes in distributed systems could cause network related timeouts. TCP/IP provides plenty of feedbacks for different kinds of network errors. If we use these feedbacks properly, we could get rid of many network timeouts.

Lifetime and mutability are simple concepts. However, when combined with reborrow and subtyping, it could get very confusing. Here's a summary of my current understandings.

Rust ownership and lifetime are very powerful tools. They were designed to help compiler get better optimization. As a side effect, they could force programmers to write cleaner code, even design better interfaces. Here are some theories and examples that may help understanding lifetimes. Disclaimer: I didn't read the compiler code so I could be wrong.

Linux bpf verifier allows only one specific pattern for accessing skb data. To help the clang compiler generate the required access pattern, we have to write C code in a certain way to reflect this pattern.

Linux bpf maps are used to share data among bpf programs and user applications. A bpf map could be created by simply declaring a bpf_elf_map struct. Under the hood, lots of things work together to set up the maps.

The file /etc/resolv.conf defines nameservers for glibc. Normally this file contains 2 or 3 nameservers for redundancy. When several nameservers serve behind one virtual IP, it is still better to have multiple (max 3) duplicated entries than having just one virtual IP entry in /etc/resolv.conf.

One challenge of TCP is to handle stale packets. Packets from previous connections may get delayed in the network and interfere with new connections. To solve this problem, TCP enforces the following rule on sequence numbers.