3 releases

0.1.2 Sep 4, 2024
0.1.1 Aug 28, 2024
0.1.0 Aug 27, 2024

#552 in Database interfaces

Download history 312/week @ 2024-08-27 138/week @ 2024-09-03 1/week @ 2024-09-10 6/week @ 2024-09-17

128 downloads per month

Apache-2.0

17KB
218 lines

Warning

This library is experimental; both the API and ABI are subject to change.

Description

This library maintains a thread-local mapping of keys to values. Each key and value is an arbitrary byte array.

The core goal of the design is that the map for a thread may be validly read from that thread whenever user code is stopped; for example, in a signal handler, a debugger, or an eBPF program. This should work even if the thread happens to be suspended in the middle of one of the functions of this library.

The intended purpose is to store custom labels for annotating stack traces during profiling; for example, client code might set the label customer_id whenever it is processing a request for a particular customer, and a CPU profiler might then record that value whenever it interrupts the program to collect a stack trace.

The library exposes a C API (in customlabels.h), a Rust API documented here, and an ABI for reading by external code (e.g., profilers or debuggers).

Supported Configurations

Language: any language that can link against C code.

Platform: Linux on x86-64 or aarch64 (64-bit ARM).

Using from Rust

Simply depend on the custom-labels crate.

Using from C or C++ (shared library)

For a release build:

CFLAGS="-O2" make

For a debug build:

CFLAGS="-O0 -g" make

Either will produce a library called libcustomlabels.so in the repository root, which should be linked against during your build process.

Using from C or C++ (main executable)

Ensure that customlabels.c is linked into your executable and that customlabels.h is available in the include path for any source file from which you want to use custom labels. The details of this will depend on your build system.

ABI

For profiler authors, the ABI is v0 of the Custom Labels ABI described here.

Acknowledgements

  • The approach was partially influenced by the APM/universal profiling integration described here.

Dependencies

~0–2.2MB
~44K SLoC