gerhobbelt / quill

Asynchronous Low Latency C++ Logging Library

Quill

Asynchronous Low Latency C++ Logging Library

• Introduction
• Documentation
• Features
• Performance
• Basic Usage
• CMake Integration
• Design
• License

homebrew	vcpkg	conan
brew install quill	vcpkg install quill	quill/[>=1.2.3]

Introduction

Quill is a cross-platform low latency logging library based on C++14/C++17.

There are two versions on the library:

v1.7 : C++14 - Bug fix only

v2 : C++17 - New features and actively maintained

Documentation

ReadtheDocs

The examples folder is also a good source of documentation.

Features

• Low latency logging See Benchmarks.
• Format outside the hot-path in a backend logging thread. For non-built-in types ostream::operator<<() is called on a copy of the object by the backend logging thread. Unsafe to copy non-trivial user defined are detected in compile time. Those types can be tagged as safe-to-copy to avoid formatting them on the hot path. See User Defined Types.
• Custom formatters. Logs can be formatted based on a user specified pattern. See Formatters.
• Support for rdtsc, chrono or custom clock (usefull for simulations) for timestamp generation.
• Support for log stack traces. Store log messages in a ring buffer and display later on a higher severity log statement or on demand. See Backtrace Logging.
• Various logging targets. See Handlers.
  • Console logging with colours support.
  • File Logging
  • Rotating log files
  • Time rotating log files
  • JSON logging
  • Custom Handlers
• Filters for filtering log messages. See Filters.
• Ability to produce JSON structured log. See Structured-Log
• guaranteed non-blocking or non-guaranteed logging. In non-guaranteed mode there is no heap allocation of a new queue but log messages can be dropped. See FAQ.
• Support for wide character logging and wide character filenames (Windows and v1.7.x only).
• Log statements in timestamp order even when produced by different threads. This makes debugging multithreading applications easier.
• Log levels can be completely stripped out at compile time reducing if branches.
• Clean warning-free codebase even on high warning levels.
• Crash safe behaviour with a built-in signal handler.
• Type safe python style API with compile type checks and built-in support for logging STL types/containers by using the excellent {fmt} library.

Performance

Latency

Log Numbers

The following message is logged 100'000 times per thread LOG_INFO(logger, "Logging int: {}, int: {}, double: {}", i, j, d).

The results in the tables below are in nanoseconds (ns).

1 Thread

Library	50th	75th	90th	95th	99th	99.9th
Quill v2.8.0 Unbounded Queue	20	21	24	25	27	34
Quill v2.8.0 Bounded Queue	17	19	21	22	26	36
fmtlog	16	19	21	22	27	40
MS BinLog	41	43	44	46	66	118
PlatformLab NanoLog	53	66	75	80	92	106
Reckless	62	75	79	84	94	103
Iyengar NanoLog	164	186	213	232	305	389
spdlog	694	761	838	887	996	1143
g3log	5398	5639	5875	6025	6327	6691

4 Threads

Library	50th	75th	90th	95th	99th	99.9th
Quill v2.8.0 Unbounded Queue	20	22	24	26	28	35
Quill v2.8.0 Bounded Queue	17	19	21	22	26	36
fmtlog	16	19	21	23	26	35
MS BinLog	42	44	46	48	76	118
PlatformLab NanoLog	56	67	77	82	95	159
Reckless	46	62	78	92	113	155
Iyengar NanoLog	150	168	247	289	355	456
spdlog	728	828	907	959	1140	1424
g3log	5103	5318	5525	5657	5927	6279

Log Numbers and Large Strings

The following message is logged 100'000 times per thread LOG_INFO(logger, "Logging int: {}, int: {}, string: {}", i, j, large_string). The large string is over 35 characters to avoid short string optimisation of std::string

1 Thread

Library	50th	75th	90th	95th	99th	99.9th
Quill v2.8.0 Unbounded Queue	31	33	35	36	39	48
Quill v2.8.0 Bounded Queue	30	32	33	35	43	51
fmtlog	29	31	34	37	44	53
MS BinLog	50	51	53	56	77	127
PlatformLab NanoLog	71	86	105	117	136	158
Reckless	215	242	268	284	314	517
Iyengar NanoLog	172	191	218	238	312	401
spdlog	653	708	770	831	950	1083
g3log	4802	4998	5182	5299	5535	5825

4 Threads

Library	50th	75th	90th	95th	99th	99.9th
Quill v2.8.0 Unbounded Queue	31	33	35	37	40	48
Quill v2.8.0 Bounded Queue	29	31	33	35	41	49
fmtlog	29	31	35	37	44	54
MS BinLog	50	52	54	58	86	130
PlatformLab NanoLog	69	82	99	111	134	194
Reckless	187	209	232	247	291	562
Iyengar NanoLog	159	173	242	282	351	472
spdlog	679	751	839	906	1132	1478
g3log	4739	4955	5157	5284	5545	5898

The benchmarks are done on Ubuntu - Intel(R) Xeon(R) Gold 6254 CPU @ 3.10GHz with GCC 12.2

Each thread is pinned on a different cpu. Unfortunately the cores are not isolated on this system. If the backend logging thread is run in the same CPU as the caller hot-path threads, that slows down the log message processing on the backend logging thread and will cause the SPSC queue to fill faster and re-allocate.

Continuously logging messages in a loop makes the consumer (backend logging thread) unable to follow up and the queue will have to re-allocate or block for most logging libraries expect very high throughput binary loggers like PlatformLab Nanolog.

Therefore, a different approach was followed that suits more to a real time application:

1. 20 messages are logged in a loop.
2. calculate/store the average latency for those messages.
3. wait between 1-2 ms.
4. repeat for n iterations.

I run each logger benchmark 4 times and the above latencies are the second-best result.

The benchmark code and results can be found here.

Throughput

The main focus of the library is not throughput. The backend logging thread is a single thread responsible for formatting, ordering the log messages from multiple hot threads and finally outputting everything as human-readable text. The logging thread always empties all the queues of the hot threads on the highest priority (to avoid allocating a new queue or dropping messages on the hot path). To achieve that, it internally buffers the log messages and then writes them later when the hot thread queues are empty or when a limit is reached backend_thread_transit_events_soft_limit.

I haven't found an easy way to compare the throughput against other logging libraries while doing asynchronous logging. For example some libraries will drop the log messages ending in producing much smaller log files than the expected, other libraries only offer an async flush meaning that you never really know when the logging thread has finished processing everything.

Quill has a blocking flush() guaranteeing every single log message from the hot threads up to that point is flushed to the file. The maximum throughput is measured as the max log messages number the backend logging thread can write to the file per second.

Benchmark code can be found here

Measured on the same system as the latency benchmarks above for 4 million messages produces a log file of 476 mb

1.76 million msgs/sec average, total time elapsed 2266 ms, total log messages 4 million

Basic usage

#include "quill/Quill.h"

int main()
{
  quill::Config cfg;
  cfg.enable_console_colours = true;
  quill::configure(cfg);
  quill::start();

  quill::Logger* logger = quill::get_logger();
  logger->set_log_level(quill::LogLevel::TraceL3);

  // enable a backtrace that will get flushed when we log CRITICAL
  logger->init_backtrace(2, quill::LogLevel::Critical);

  LOG_BACKTRACE(logger, "Backtrace log {}", 1);
  LOG_BACKTRACE(logger, "Backtrace log {}", 2);

  LOG_INFO(logger, "Welcome to Quill!");
  LOG_ERROR(logger, "An error message. error code {}", 123);
  LOG_WARNING(logger, "A warning message.");
  LOG_CRITICAL(logger, "A critical error.");
  LOG_DEBUG(logger, "Debugging foo {}", 1234);
  LOG_TRACE_L1(logger, "{:>30}", "right aligned");
  LOG_TRACE_L2(logger, "Positional arguments are {1} {0} ", "too", "supported");
  LOG_TRACE_L3(logger, "Support for floats {:03.2f}", 1.23456);
}

Output

CMake-Integration

External

Building and Installing Quill as Static Library

git clone http://github.com/odygrd/quill.git
mkdir cmake_build
cd cmake_build
make install

Note: To install in custom directory invoke cmake with -DCMAKE_INSTALL_PREFIX=/quill/install-dir/

Building and Installing Quill as Static Library With External libfmt

cmake -DCMAKE_PREFIX_PATH=/my/fmt/fmt-config.cmake-directory/ -DQUILL_FMT_EXTERNAL=ON -DCMAKE_INSTALL_PREFIX=/quill/install-dir/'

Then use the library from a CMake project, you can locate it directly with find_package()

Directory Structure

my_project/
├── CMakeLists.txt
├── main.cpp

CMakeLists.txt

# Set only if needed - quill was installed under a custom non-standard directory
set(CMAKE_PREFIX_PATH /test_quill/usr/local/)

find_package(quill REQUIRED)

# Linking your project against quill
add_executable(example main.cpp)
target_link_libraries(example PRIVATE quill::quill)

main.cpp

See basic usage

Embedded

To embed the library directly, copy the source folder to your project and call add_subdirectory() in your CMakeLists.txt file

Directory Structure

my_project/
├── quill/            (source folder)
├── CMakeLists.txt
├── main.cpp

CMakeLists.txt

cmake_minimum_required(VERSION 3.1.0)
project(my_project)

set(CMAKE_CXX_STANDARD 14)

add_subdirectory(quill)

add_executable(my_project main.cpp)
target_link_libraries(my_project PRIVATE quill::quill)

main.cpp

See basic usage

Design

License

Quill is licensed under the MIT License

Quill depends on third party libraries with separate copyright notices and license terms. Your use of the source code for these subcomponents is subject to the terms and conditions of the following licenses.

• (MIT License) {fmt} (http://github.com/fmtlib/fmt/blob/master/LICENSE.rst)
• (MIT License) doctest (http://github.com/onqtam/doctest/blob/master/LICENSE.txt)