histlog

Provides off-thread serialization of HdrHistogram interval logs to file.

For use with the hdrhistogram crate, a rust port of Gil Tene's HdrHistogram, that provides a clean interface with sane defaults for off-thread serialization of HdrHistogram interval logs to file.

purpose

HdrHistogram is often used to measure latency. Generally, if something is important enough to measure latency, it's unlikely you want to write to a file on the same thread.

One option would be to serialize to an in-memory buffer (e.g. Vec<u8>). However, this would still require allocating to the buffer, and would eventually require a lot of memory for a long-running process.

histlog::HistLog allows the hot thread to pass off it's hdrhistogram::Histogram at regular intervals to a designated writer thread that can afford to dilly dally with IO. The interval log is written incrementally and can be inspected and analyzed while the program is still running.

HistLog relies completely on the hdrhistogram crate, both for the in-memory recording of values and serialization. What it does provide is off-thread writing with a clean interface and sane defaults that make it relatively easy to use.

examples

A HistLog has a "series" name and a "tag." The HdrHistogram interval log format provides for one tag per entry. The series name is used to name the file the interval log is written to:

use std::time::*;

let log_dir = "/tmp/path/to/logs";
let series = "server-latency";          // used to name the log file
let tag = "xeon-e7-8891-v2";            // recorded with each entry
let freq = Duration::from_secs(1);      // how often results sent to writer thread

// `HistLog::new` could fail creating file, `hdrhistogram::Histogram`
let mut server1 = histlog::HistLog::new(log_dir, series, tag, freq).unwrap();

// use `HistLog::clone_with_tag` to serialize a separate tag to same file.
let mut server2 = server1.clone_with_tag("xeon-e5-2670");

for i in 0..1000u64 { // dummy data
    server1.record(i).unwrap(); // call to `hdrhistogram::Histogram::record` could fail
    server2.record(i * 2).unwrap();
}

assert_eq!(server1.path(), server2.path()); // both being saved to same file, via same writer thread

HistLog's api design is built for event loops. Each iteration of the loop, new values are recorded, and the current time is checked to see whether the current Histogram should be passed off to the writer thread:

use std::time::*;

let mut spintime = histlog::HistLog::new("/tmp/var/hist", "spintime", "main", Duration::from_secs(60)).unwrap();

let mut loop_time = Instant::now();
let mut prev: Instant;

loop {
    prev = loop_time;
    loop_time = Instant::now();
    spintime.record(histlog::nanos(loop_time - prev)).unwrap(); // nanos: Duration -> u64
    spintime.check_send(loop_time); // sends to writer thread if elapsed > freq,
    // or...
    spintime.check_try_send(loop_time).unwrap(); // non-blocking equivalent (can fail)

    // do important stuff ...
}

logs

Logs are saved to <log dir>/<series name>.<datetime>.hdrhistogram-interval-log.v2.gz.

Format of log is like this:

#[StartTime: 1544631293.283 (seconds since epoch)]
#[BaseTime: 0.000 (seconds since epoch)]
Tag=xeon-e7-8891-v2,1544631293.283,0.003,999.000,HISTFAAAAC94Ae3GMRUAMAgD0bRI6FovNVcHmGREAgNR [...]
Tag=xeon-e5-2670,1544631293.283,0.003,999.000,HISTFAAAABx4AZNpmSzMwMDAxAABzFCaEUoz2X+AsQA/awK [...]
[...]

Only the histogram data is compressed (deflate), so a .gz extension is perhaps misleading.

Log file can be viewed/analyzed here (javascript, runs locally) or with the Java-based HistogramLogAnalyzer.

Full documentation of log serialization available from the hdrhistogram crate.

features

• minstant: replaces std::time::Instant with minstant::Instant, which is generally faster and a drop-in replacement with the same API.
• smol_str: switches smol_str::SmolStr for &'static str as the type for both series names (SeriesName type alias) and tags (Tag type alias). SmolStr allows dynamic strings without heap allocations up to a certain size, so this feature enables dynamic values for both series names and tag values.

limitations

• HistLog::check_send and HistLog::check_try_send create a new hdrhistogram::Histogram and send the current/prev one to the writer thread each interval. Internally, an hdrhistogram::Histogram uses a Vec to store its counts, so there's an allocation involved.
• Only u64 values can be recorded (see the C type alias).