The zpool_influxdb program produces
influxdb line protocol
compatible metrics from zpools. In the UNIX tradition, zpool_influxdb
does one thing: read statistics from a pool and print them to
stdout. In many ways, this is a metrics-friendly output of
statistics normally observed via the zpool
command.
There are many implementations of ZFS on many OSes. The current version is tested to work on:
- ZFSonLinux version 0.7 and later
- cstor for userland ZFS (uZFS)
This should compile and run on other ZFS versions, though many do not have the latency histograms. Pull requests are welcome.
When run without arguments, zpool_influxdb runs once, reading data from all imported pools, and prints to stdout.
zpool_influxdb [options] [poolname]
If no poolname is specified, then all pools are sampled.
option | short option | description |
---|---|---|
--execd | -e | For use with telegraf's execd plugin. When [enter] is pressed, the pools are sampled. To exit, use [ctrl+D] |
--no-histogram | -n | Do not print histogram information |
--sum-histogram-buckets | -s | Sum histogram bucket values |
--help | -h | Print a short usage message |
The histogram data collected by ZFS is stored as independent bucket values. This works well out-of-the-box with an influxdb data source and grafana's heatmap visualization. The influxdb query for a grafana heatmap visualization looks like:
field(disk_read) last() non_negative_derivative(1s)
Another method for storing histogram data sums the values for lower-value
buckets. For example, a latency bucket tagged "le=10" includes the values
in the bucket "le=1".
This method is often used for prometheus histograms.
The zpool_influxdb --sum-histogram-buckets
option presents the data from ZFS
as summed values.
The following measurements are collected:
measurement | description | zpool equivalent |
---|---|---|
zpool_stats | general size and data | zpool list |
zpool_scan_stats | scrub, rebuild, and resilver statistics (omitted if no scan has been requested) | zpool status |
zpool_vdev_stats | per-vdev statistics | zpool iostat -q |
zpool_io_size | per-vdev I/O size histogram | zpool iostat -r |
zpool_latency | per-vdev I/O latency histogram | zpool iostat -w |
zpool_vdev_queue | per-vdev instantaneous queue depth | zpool iostat -q |
zpool_stats contains top-level summary statistics for the pool. Performance counters measure the I/Os to the pool's devices.
label | description |
---|---|
name | pool name |
state | pool state, as shown by zpool status |
field | units | description |
---|---|---|
alloc | bytes | allocated space |
free | bytes | unallocated space |
size | bytes | total pool size |
read_bytes | bytes | bytes read since pool import |
read_errors | count | number of read errors |
read_ops | count | number of read operations |
write_bytes | bytes | bytes written since pool import |
write_errors | count | number of write errors |
write_ops | count | number of write operations |
Once a pool has been scrubbed, resilvered, or rebuilt, the zpool_scan_stats contain information about the status and performance of the operation. Otherwise, the zpool_scan_stats do not exist in the kernel, and therefore cannot be reported by this collector.
label | description |
---|---|
name | pool name |
function | name of the scan function running or recently completed |
state | scan state, as shown by zpool status |
field | units | description |
---|---|---|
errors | count | number of errors encountered by scan |
examined | bytes | total data examined during scan |
to_examine | bytes | prediction of total bytes to be scanned |
pass_examined | bytes | data examined during current scan pass |
processed | bytes | data reconstructed during scan |
to_process | bytes | total bytes to be repaired |
rate | bytes/sec | examination rate |
start_ts | epoch timestamp | start timestamp for scan |
pause_ts | epoch timestamp | timestamp for a scan pause request |
end_ts | epoch timestamp | completion timestamp for scan |
paused_t | seconds | elapsed time while paused |
remaining_t | seconds | estimate of time remaining for scan |
The ZFS I/O (ZIO) scheduler uses five queues to schedule I/Os to each vdev. These queues are further divided into active and pending states. An I/O is pending prior to being issued to the vdev. An active I/O has been issued to the vdev. The scheduler and its tunable parameters are described at the ZFS on Linux wiki. The ZIO scheduler reports the queue depths as gauges where the value represents an instantaneous snapshot of the queue depth at the sample time. Therefore, it is not unusual to see all zeroes for an idle pool.
label | description |
---|---|
name | pool name |
vdev | vdev name (root = entire pool) |
field | units | description |
---|---|---|
sync_r_active_queue | entries | synchronous read active queue depth |
sync_w_active_queue | entries | synchronous write active queue depth |
async_r_active_queue | entries | asynchronous read active queue depth |
async_w_active_queue | entries | asynchronous write active queue depth |
async_scrub_active_queue | entries | asynchronous scrub active queue depth |
sync_r_pend_queue | entries | synchronous read pending queue depth |
sync_w_pend_queue | entries | synchronous write pending queue depth |
async_r_pend_queue | entries | asynchronous read pending queue depth |
async_w_pend_queue | entries | asynchronous write pending queue depth |
async_scrub_pend_queue | entries | asynchronous scrub pending queue depth |
ZFS tracks the latency of each I/O in the ZIO pipeline. This latency can be useful for observing latency-related issues that are not easily observed using the averaged latency statistics.
The histogram fields show cumulative values from lowest to highest. The largest bucket is tagged "le=+Inf", representing the total count of I/Os by type and vdev.
label | description |
---|---|
le | bucket for histogram, latency is less than or equal to bucket value in seconds |
name | pool name |
path | for leaf vdevs, the device path name, otherwise omitted |
vdev | vdev name (root = entire pool) |
field | units | description |
---|---|---|
total_read | operations | read operations of all types |
total_write | operations | write operations of all types |
disk_read | operations | disk read operations |
disk_write | operations | disk write operations |
sync_read | operations | ZIO sync reads |
sync_write | operations | ZIO sync writes |
async_read | operations | ZIO async reads |
async_write | operations | ZIO async writes |
scrub | operations | ZIO scrub/scan reads |
trim | operations | ZIO trim (aka unmap) writes |
ZFS tracks I/O throughout the ZIO pipeline. The size of each I/O is used to create a histogram of the size by I/O type and vdev. For example, a 4KiB write to mirrored pool will show a 4KiB write to the top-level vdev (root) and a 4KiB write to each of the mirror leaf vdevs.
The ZIO pipeline can aggregate I/O operations. For example, a contiguous series of writes can be aggregated into a single, larger I/O to the leaf vdev. The independent I/O operations reflect the logical operations and the aggregated I/O operations reflect the physical operations.
The histogram fields show cumulative values from lowest to highest. The largest bucket is tagged "le=+Inf", representing the total count of I/Os by type and vdev.
Note: trim I/Os can be larger than 16MiB, but the larger sizes are accounted in the 16MiB bucket.
label | description |
---|---|
le | bucket for histogram, I/O size is less than or equal to bucket value in bytes |
name | pool name |
path | for leaf vdevs, the device path name, otherwise omitted |
vdev | vdev name (root = entire pool) |
field | units | description |
---|---|---|
sync_read_ind | blocks | independent sync reads |
sync_write_ind | blocks | independent sync writes |
async_read_ind | blocks | independent async reads |
async_write_ind | blocks | independent async writes |
scrub_read_ind | blocks | independent scrub/scan reads |
trim_write_ind | blocks | independent trim (aka unmap) writes |
sync_read_agg | blocks | aggregated sync reads |
sync_write_agg | blocks | aggregated sync writes |
async_read_agg | blocks | aggregated async reads |
async_write_agg | blocks | aggregated async writes |
scrub_read_agg | blocks | aggregated scrub/scan reads |
trim_write_agg | blocks | aggregated trim (aka unmap) writes |
Telegraf v1.6.2 and later support unsigned 64-bit integers which more
closely matches the uint64_t values used by ZFS. By default, zpool_influxdb
will mask ZFS' uint64_t values and use influxdb line protocol integer type.
Eventually the monitoring world will catch up to the times and support
unsigned integers. To support unsigned, define SUPPORT_UINT64 and compile
as described in CMakeLists.txt
Building is simplified by using cmake.
It is as simple as possible, but no simpler.
By default, ZFSonLinux
installs the necessary header and library files in /usr/local.
If you place those files elsewhere, either edit CMakeLists.txt and
change the ZFS_INSTALL_BASE or pass it with -D ZFS_INSTALL_BASE=/usr
on the cmake command line:
cmake .
make
If successful, the zpool_influxdb executable is created.
Installation is left as an exercise for the reader because there are many different methods that can be used. Ultimately the method depends on how the local metrics collection is implemented and the local access policies.
To install the zpool_influxdb executable in INSTALL_DIR, use
make install
The simplest method is to use the exec agent in telegraf. For convenience, a sample config file is zpool_influxdb.conf which can be placed in the telegraf config-directory (often /etc/telegraf/telegraf.d). Telegraf can be restarted to read the config-directory files.
- Like the zpool command, zpool_influxdb takes a reader
lock on spa_config for each imported pool. If this lock blocks,
then the command will also block indefinitely and might be
unkillable. This is not a normal condition, but can occur if
there are bugs in the kernel modules.
For this reason, care should be taken:
- avoid spawning many of these commands hoping that one might finish
- avoid frequent updates or short sample time intervals, because the locks can interfere with the performance of other instances of zpool or zpool_influxdb
There are a few other collectors for zpool statistics roaming around
the Internet. Many attempt to screen-scrape zpool
output in various
ways. The screen-scrape method works poorly for zpool
output because
of its human-friendly nature. Also, they suffer from the same caveats
as this implementation. This implementation is optimized for directly
collecting the metrics and is much more efficient than the screen-scrapers.
Pull requests and issues are greatly appreciated. Visit https://github.com/richardelling/zpool_influxdb