pub struct HistLog { /* private fields */ }
Expand description
Provides 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
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 rust port of HdrHistogram
, 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.
Limitations
- The series name and tags are currently limited to
&'static str
because the overhead of usingString
is prohibitive. This may change in future versions if a performant means of allowing dynamic tags presents itself that’s not inordinately complicated to use. HistLog::check_send
andHistLog::check_try_send
create a newhdrhistogram::Histogram
and send the current/prev one to the writer thread each interval. Internally, anhdrhistogram::Histogram
uses aVec
to store its counts, so there’s an allocation involved.- Only
u64
values can be recorded, currently.
Implementations
sourceimpl HistLog
impl HistLog
sourcepub fn new<P>(
save_dir: P,
series: &'static str,
tag: &'static str,
freq: Duration
) -> Result<Self, Error>where
P: AsRef<Path>,
pub fn new<P>(
save_dir: P,
series: &'static str,
tag: &'static str,
freq: Duration
) -> Result<Self, Error>where
P: AsRef<Path>,
Create a new HistLog
.
If save_dir
does not exist, will attempt to create it (which could
fail). Creating a new log file could fail. Spawning the writer thread could fail.
pub fn new_with_tag(&self, tag: &'static str) -> Result<Self, Error>
sourcepub fn clone_with_tag(&self, tag: &'static str) -> Self
pub fn clone_with_tag(&self, tag: &'static str) -> Self
Record a new histogram with a tag
that will serialize to the
same interval log file as its parent. Each cloned HistLog
’s entries
will be written to their own lines in the log file, identifiable by tag.
Limitations
No effort is made to check whether tag
is a duplicate of a previous tag,
and using a duplicate may produce unexpected results.
pub fn clone_with_tag_and_freq(&self, tag: &'static str, freq: Duration) -> Self
sourcepub fn record(&mut self, value: u64) -> Result<(), Error>
pub fn record(&mut self, value: u64) -> Result<(), Error>
Record a single value to the histogram. This could fail if the value
is outside of the highest range permitted. See the
hdrhistogram
docs
for further deails. The hdrhistogram::Histogram
used by HistLog
is created with a significant figure of 3 (histlog::SIG_FIG
const).
sourcepub fn reset(&mut self)
pub fn reset(&mut self)
Reset the state of the internal histogram and the last sent value.
One situation this might be used is if there was a pause in recording.
sourcepub fn check_send(&mut self, loop_time: Instant) -> bool
pub fn check_send(&mut self, loop_time: Instant) -> bool
Send the current histogram to the writer thread if the elapsed time since the last send is greater than the interval frequency.
If the channel is disconnected, this will fail silently, instead of panicking.
Trait Implementations
sourceimpl Drop for HistLog
impl Drop for HistLog
sourcefn drop(&mut self)
fn drop(&mut self)
Checks if the current instance is the last remaining instance with a reference to the underlying writer thread, and, if so, sends a terminate signal to the writer thread and attempts to join it.
May Pause Up To 5ms
In the event the channel to the writer thread is full, will continue trying
to send a terminate command (busy polling the channel) until DROP_DEADLINE
has expired (currently 5ms), upon which it will abort.
If channel is disconnected, will simply abort without trying to join the writer thread.