Skip to content

Commit

Permalink
Merge pull request #167 from prometheus/improved_documentation
Browse files Browse the repository at this point in the history
Improve documentation for DirectFileStore
  • Loading branch information
Daniel Magliola committed Nov 1, 2019
2 parents fd26a66 + 81222d3 commit 367cbff
Show file tree
Hide file tree
Showing 2 changed files with 55 additions and 31 deletions.
74 changes: 47 additions & 27 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -271,7 +271,7 @@ is stored in a global Data Store object, rather than in the metric objects thems
(This "storage" is ephemeral, generally in-memory, it's not "long-term storage")

The main reason to do this is that different applications may have different requirements
for their metrics storage. Application running in pre-fork servers (like Unicorn, for
for their metrics storage. Applications running in pre-fork servers (like Unicorn, for
example), require a shared store between all the processes, to be able to report coherent
numbers. At the same time, other applications may not have this requirement but be very
sensitive to performance, and would prefer instead a simpler, faster store.
Expand Down Expand Up @@ -311,7 +311,7 @@ whether you want to report the `SUM`, `MAX` or `MIN` value observed across all p
For almost all other cases, you'd leave the default (`SUM`). More on this on the
*Aggregation* section below.

Other custom stores may also accept extra parameters besides `:aggregation`. See the
Custom stores may also accept extra parameters besides `:aggregation`. See the
documentation of each store for more details.

### Built-in stores
Expand All @@ -326,26 +326,46 @@ There are 3 built-in stores, with different trade-offs:
it's absolutely not thread safe.
- **DirectFileStore**: Stores data in binary files, one file per process and per metric.
This is generally the recommended store to use with pre-fork servers and other
"multi-process" scenarios.

Each metric gets a file for each process, and manages its contents by storing keys and
binary floats next to them, and updating the offsets of those Floats directly. When
exporting metrics, it will find all the files that apply to each metric, read them,
and aggregate them.

In order to do this, each Metric needs an `:aggregation` setting, specifying how
to aggregate the multiple possible values we can get for each labelset. By default,
they are `SUM`med, which is what most use-cases call for (counters and histograms,
for example). However, for Gauges, it's possible to set `MAX` or `MIN` as aggregation,
to get the highest/lowest value of all the processes / threads.

Even though this store saves data on disk, it's still much faster than would probably be
expected, because the files are never actually `fsync`ed, so the store never blocks
while waiting for disk. The kernel's page cache is incredibly efficient in this regard.

If in doubt, check the benchmark scripts described in the documentation for creating
your own stores and run them in your particular runtime environment to make sure this
provides adequate performance.
"multi-process" scenarios. There are some important caveats to using this store, so
please read on the section below.

### `DirectFileStore` caveats and things to keep in mind

Each metric gets a file for each process, and manages its contents by storing keys and
binary floats next to them, and updating the offsets of those Floats directly. When
exporting metrics, it will find all the files that apply to each metric, read them,
and aggregate them.

**Aggregation of metrics**: Since there will be several files per metrics (one per process),
these need to be aggregated to present a coherent view to Prometheus. Depending on your
use case, you may need to control how this works. When using this store,
each Metric allows you to specify an `:aggregation` setting, defining how
to aggregate the multiple possible values we can get for each labelset. By default,
Counters, Histograms and Summaries are `SUM`med, and Gauges report all their values (one
for each process), tagged with a `pid` label. You can also select `SUM`, `MAX` or `MIN`
for your gauges, depending on your use case.

**Memory Usage**: When scraped by Prometheus, this store will read all these files, get all
the values and aggregate them. We have notice this can have a noticeable effect on memory
usage for your app. We recommend you test this in a realistic usage scenario to make sure
you won't hit any memory limits your app may have.

**Resetting your metrics on each run**: You should also make sure that the directory where
you store your metric files (specified when initializing the `DirectFileStore`) is emptied
when your app starts. Otherwise, each app run will continue exporting the metrics from the
previous run.

**Large numbers of files**: Because there is an individual file per metric and per process
(which is done to optimize for observation performance), you may end up with a large number
of files. We don't currently have a solution for this problem, but we're working on it.

**Performance**: Even though this store saves data on disk, it's still much faster than
would probably be expected, because the files are never actually `fsync`ed, so the store
never blocks while waiting for disk. The kernel's page cache is incredibly efficient in
this regard. If in doubt, check the benchmark scripts described in the documentation for
creating your own stores and run them in your particular runtime environment to make sure
this provides adequate performance.


### Building your own store, and stores other than the built-in ones.

Expand All @@ -364,16 +384,16 @@ If you are in a multi-process environment (such as pre-fork servers like Unicorn
process will probably keep their own counters, which need to be aggregated when receiving
a Prometheus scrape, to report coherent total numbers.

For Counters and Histograms (and quantile-less Summaries), this is simply a matter of
For Counters, Histograms and quantile-less Summaries this is simply a matter of
summing the values of each process.

For Gauges, however, this may not be the right thing to do, depending on what they're
measuring. You might want to take the maximum or minimum value observed in any process,
rather than the sum of all of them. You may also want to export each process's individual
value.
rather than the sum of all of them. By default, we export each process's individual
value, with a `pid` label identifying each one.

In those cases, you should use the `store_settings` parameter when registering the
metric, to specify an `:aggregation` setting.
If these defaults don't work for your use case, you should use the `store_settings`
parameter when registering the metric, to specify an `:aggregation` setting.

```ruby
free_disk_space = registry.gauge(:free_disk_space_bytes,
Expand Down
12 changes: 8 additions & 4 deletions lib/prometheus/client/data_stores/direct_file_store.rb
Original file line number Diff line number Diff line change
Expand Up @@ -18,10 +18,14 @@ module DataStores
#
# In order to do this, each Metric needs an `:aggregation` setting, specifying how
# to aggregate the multiple possible values we can get for each labelset. By default,
# they are `SUM`med, which is what most use cases call for (counters and histograms,
# for example).
# However, for Gauges, it's possible to set `MAX` or `MIN` as aggregation, to get
# the highest value of all the processes / threads.
# Counters, Histograms and Summaries get `SUM`med, and Gauges will report `ALL`
# values, tagging each one with a `pid` label.
# For Gauges, it's also possible to set `SUM`, MAX` or `MIN` as aggregation, to get
# the highest / lowest value / or the sum of all the processes / threads.
#
# Before using this Store, please read the "`DirectFileStore` caveats and things to
# keep in mind" section of the main README in this repository. It includes a number
# of important things to keep in mind.

class DirectFileStore
class InvalidStoreSettingsError < StandardError; end
Expand Down

0 comments on commit 367cbff

Please sign in to comment.