Oban Met
Met is a distributed, compacting, multidimensional, telemetry-powered time series datastore for Oban that requires no configuration. It gathers data for queues, job counts, execution metrics, active crontabs, historic metrics, and more.
Met powers the charts and runtime details shown in the Oban Web dashboard.
Features
-
🤖 Autonomous - Supervises a collection of autonomous modules that dynamically start and stop alongside Oban instances without any code changes.
-
🎩 Distributed - Metrics are shared between all connected nodes via pubsub. Leadership is used to restrict expensive operations, such as performing counts, to a single node.
-
📼 Recorded - Telemetry events and scraped data are stored in-memory as time series data. Values are stored as either gauges or space efficient "sketches".
-
🪐 Multidimensional - Metrics are stored with labels such as
node,queue,worker, etc. that can be filtered and grouped dynamically at runtime. -
🗜️ Compacting - Time series values are periodically compacted into larger windows of time to save space and optimize querying historic data. Compaction periods use safe defaults, but are configurable.
-
✏️ Estimating - In supporting systems (Postgres), count queries use optimized estimates automatically for tables with a large number of jobs.
-
🔎 Queryable - Historic metrics may be filtered and grouped by any label, sliced by arbitrary time intervals, and numeric values aggregated at dynamic percentiles (e.g. P50, P99) without pre-computed histogram buckets.
-
🤝 Handoff - Ephemeral data storage via data replication with handoff between nodes. All nodes have a shared view of the cluster's data and, with the
Oban.Notifiers.PGnotifier, new nodes are caught up when they come online.
Installation
Oban Met is included with Oban Web and manual installation is only necessary in hybrid environments (separate Web and Worker nodes).
To receive metrics from non-web nodes in a system with separate "web" and "worker" applications
you must explicitly include oban_met as a dependency for "workers".
{:oban_met, "~> 1.3"}
Met requires Oban v2.23 or later.
Usage
No configuration is necessary and Oban Met will start automatically in a typical application. A variety of options are provided for more complex or nuanced usage.
Querying Metrics
Met records four series, each labeled by node, queue, and state, and the job series also by
worker:
exec_count- a gauge of jobs that finished executing each second, labeled by their state after execution, e.g.completedorretryableexec_time- a sketch of job execution timeswait_time- a sketch of how long jobs waited in the queue before executingfull_count- a gauge of jobs in the database, counted periodically by the leader node
Use Oban.Met.latest/3 for current values and Oban.Met.timeslice/3 for history, filtered and
grouped by any label:
Oban.Met.latest(:full_count, group: "state")
%{"available" => 9, "completed" => 80, "executing" => 5}
Oban.Met.timeslice(:exec_time, by: 5, group: "queue", operation: {:pct, 0.95})
[{11, 9970235387.03, "analysis"}, ..., {0, 1520977874.33, "events"}]
Producer checks from every node are available through Oban.Met.checks/1, and the merged crontab
from every node through Oban.Met.crontab/1.
Configuration
Met starts without configuration, but every supervised component accepts options that may be provided two ways:
-
Through the application environment, where they apply to all Oban instances:
config :oban_met, reporter: [estimate_limit: 200_000] -
Through the plugin specification, where they apply to a single instance and take precedence over the application environment:
{Oban.Met, reporter: [estimate_limit: 200_000]}
Options are namespaced by component: :cronitor, :examiner, :listener, :recorder, and
:reporter. Most are internal and best left alone, but the sections below cover the options worth
tuning.
Auto Start
Supervised Met instances start automatically along with Oban instances unless Oban is in testing mode. You can disable auto-starting globally with application configuration:
config :oban_met, auto_start: false
Then, start instances as a child directly within your Oban app's plugins:
plugins: [
Oban.Met,
...
]
By default Met only starts when Oban's :testing mode is :disabled. To run Met under a specific
testing mode, for example to inspect metrics during :manual tests, add the mode to
:auto_testing_modes:
config :oban_met, auto_testing_modes: [:disabled, :manual]
Customizing Estimates
The estimate_limit determines when counts switch from an accurate count(*) query to a much more
efficient, but less accurate, estimate function. The limit applies to each state's total across all
queues: once a state reaches the limit, its count for every queue is estimated on the following
checks. Estimates are only available on Postgres, and other databases always use exact counts.
The default limit is a conservative 50k, which may be too low for systems with insert spikes. This declares an override to set the limit to 200k:
{Oban.Met, reporter: [estimate_limit: 200_000]}
Customizing Check Interval
The reporter periodically counts jobs in each state/queue combination to provide the counts displayed in Oban Web. By default, counts are checked every 1 second. For systems where real-time counts aren't critical, or to reduce database load, you can increase the interval:
{Oban.Met, reporter: [check_interval: :timer.seconds(5)]}
The list of queues used for estimates is refreshed every 60 checks, so newly added queues may take longer to appear in estimated counts with a longer interval.
Explicit Migrations
Met will create the necessary estimate function automatically when possible. The migration isn't necessary under normal circumstances, but is provided to avoid permission issues or allow full control over database changes. If the database role Met connects with can't create or replace the function, counts fail on every check and the error is logged until the migration is used instead.
mix ecto.gen.migration add_oban_met
Open the generated migration and delegate the up/0 and down/0 functions to
Oban.Met.Migration:
defmodule MyApp.Repo.Migrations.AddObanMet do
use Ecto.Migration
def up, do: Oban.Met.Migration.up()
def down, do: Oban.Met.Migration.down()
end
Pass a :prefix to both functions, e.g. Oban.Met.Migration.up(prefix: "private"), when Oban
runs in an alternate prefix.
Then configure the reporter not to auto-migrate:
config :oban_met, reporter: [auto_migrate: false]
Customizing the Estimate Function
By default the reporter installs and calls the built-in PostgreSQL oban_count_estimate function.
On other Postgres-compatible engines that can't run that definition (e.g. CockroachDB) you can
supply your own (prefix -> ddl) builder and keep auto_migrate enabled. The returned DDL must
create a <prefix>.oban_count_estimate(state text, queue text) function returning an integer estimate:
{Oban.Met, reporter: [estimate_function: &MyApp.ObanEstimate.build/1]}
Compaction Periods
Recorded metrics are periodically compacted into coarser windows to bound memory use. Each period
is a {step, duration} tuple in seconds, where step is the resolution and duration is how long
values are retained at that resolution, after the periods before it. The defaults keep about three
and a half hours of history:
{Oban.Met, recorder: [compact_periods: [{1, 120}, {5, 900}, {30, 2_000}, {60, 9_300}]]}
Web Charts {: .neutral}
Oban Web's charts are rendered from these recorded metrics. Drastically shortening the periods reduces the history available to plot and may impact chart rendering.
Sketch Time Unit
Timing metrics (execution time, queue time) are stored in space-efficient quantile sketches. By
default, values are recorded in :native time units (nanoseconds on most systems), which provides
maximum precision but uses more space.
For reduced storage size (~20% smaller) and better bin consolidation, you can configure sketches to store timing in milliseconds:
config :oban_met, sketch_time_unit: :millisecond
This is a compile-time option that affects how timing values are binned in sketches, so changing it
requires recompiling Met, e.g. mix deps.compile oban_met --force.
Cluster Consistency {: .warning}
All nodes in a cluster must use the same
sketch_time_unitsetting. Mixing units between nodes will produce incorrect aggregated metrics.After changing this setting, existing metrics will be invalid until compaction ages them out (about three and a half hours with default compaction periods).
Contributing
To run the test suite you must have PostgreSQL 12+. Once dependencies are installed, setup the databases and run necessary migrations:
mix test.setup
Use mix test.reset to drop, recreate, and migrate the test databases.
Community
There are a few places to connect and communicate with other Oban users:
- Ask questions and discuss #oban on the Elixir Forum
- Request an invitation and join the #oban channel on Slack
- Learn about bug reports and upcoming features in the issue tracker