Inter process events for Nginx worker processes
- Name
- Status
- Synopsis
- Description
- Troubleshooting
- Methods
- Installation
- Bugs and Patches
- Author
- Copyright and License
- History
- See Also
This library is production ready.
http {
lua_package_path "/path/to/lua-resty-worker-events/lib/?.lua;;";
# the size depends on the number of event to handle:
lua_shared_dict process_events 1m;
init_worker_by_lua_block {
local ev = require "resty.worker.events"
local handler = function(data, event, source, pid)
print("received event; source=",source,
", event=",event,
", data=", tostring(data),
", from process ",pid)
end
ev.register(handler)
local ok, err = ev.configure {
shm = "process_events", -- defined by "lua_shared_dict"
timeout = 2, -- life time of unique event data in shm
interval = 1, -- poll interval (seconds)
wait_interval = 0.010, -- wait before retry fetching event data
wait_max = 0.5, -- max wait time before discarding event
shm_retries = 999, -- retries for shm fragmentation (no memory)
}
if not ok then
ngx.log(ngx.ERR, "failed to start event system: ", err)
return
end
}
server {
...
# example for polling:
location = /some/path {
default_type text/plain;
content_by_lua_block {
-- manually call `poll` to stay up to date, can be used instead,
-- or together with the timer interval. Polling is efficient,
-- so if staying up-to-date is important, this is preferred.
require("resty.worker.events").poll()
-- do regular stuff here
}
}
}
}This module provides a way to send events to the other worker processes in an Nginx server. Communication is through a shared memory zone where event data will be stored.
The order of events in all workers is guaranteed to be the same.
The worker process will setup a timer to check for events in the background. The module follows a singleton pattern and hence runs once per worker. If staying up-to-date is important though, the interval can be set to a lesser frequency and a call to poll upon each request received makes sure everything is handled as soon as possible.
The design allows for 3 usecases;
- broadcast an event to all workers processes, see post. In this case the order of the events is guaranteed to be the same in all worker processes. Example; a healthcheck running in one worker, but informing all workers of a failed upstream node.
- broadcast an event to the local worker only, see post_local.
- coalesce external events to a single action. Example; all workers watch
external events indicating an in-memory cache needs to be refreshed. When
receiving it they all post it with a unique event hash (all workers generate the
same hash), see
uniqueparameter of post. Now only 1 worker will receive the event only once, so only one worker will hit the upstream database to refresh the in-memory data.
This module itself will fire two events with source="resty-worker-events";
event="started"when the module is first configured (note: the event handler must be registered before calling configure to be able to catch the event)event="stopping"when the worker process exits (based on a timerprematuresetting)
See event_list for using events without hardcoded magic values/strings.
To properly size the shm, it is important to understand how it is being used. Event data is stored in the shm to pass it to the other workers. As such there are 2 types of entries in the shm:
- events that are to be executed by only a single worker (see the
uniqueparameter of thepostmethod). These entries get attlin the shm and will hence expire. - all other events (except local events which do not use the SHM). In these
cases there is no
ttlset.
The result of the above is that the SHM will always be full! so that is not a metric to investigate at.
How to prevent problems:
-
the SHM size must at least be a multiple of the maximum payload expected. It must be able to cater for all the events that might be send within one
interval(seeconfigure). -
no memoryerrors cannot be resolved by making the SHM bigger. The only way to resolve those is by increasing theshm_retriesoption passed toconfigure(which already has a high default). This is because the error is due to fragmentation and not a lack of memory. -
the
waiting for event data timed outerror happens if event data gets evicted before all the workers got to deal with it. This can happen if there is a burst of (large-payload) events. To resolve these:- try to avoid big event payloads
- use a smaller
interval, so workers check for (and deal with) events more frequently (seeintervaloption as passed toconfigure) - increase the SHM size, such that it can hold all the event data that might be send within 1 interval.
syntax: success, err = events.configure(opts)
Will initialize the event listener. This should typically be called from the
init_by_lua handler, because it will make sure all workers start with the
first event. In case of a reload of the system (starting new and stopping old
workers) past events will not be replayed. And because the order in which
workers reload cannot be guaranteed, also the event start cannot be guaranteed.
So if some sort of state is derived from the events you have to manage that
state separately.
The opts parameter is a Lua table with named options:
shm: (required) name of the shared memory to use. Event data will not expire, so the module relies on the shm lru mechanism to evict old events from the shm. As such the shm should probably not be used for other purposes.shm_retries: (optional) number of retries when the shm returns "no memory" on posting an event, default 999. Each time there is an insertion attempt and no memory is available (either no space is available or the memory is available but fragmented), "up to tens" of old entries are evicted. After that, if there's still no memory available, the "no memory" error is returned. Retrying the insertion triggers the eviction phase several times, increasing the memory available as well as the probability of finding a large enough contiguous memory block available for the new event data.interval: (optional) interval to poll for events (in seconds), default 1. Set to 0 to disable polling.wait_interval: (optional) interval between two tries when a new eventid is found, but the data is not available yet (due to asynchronous behaviour of the worker processes)wait_max: (optional) max time to wait for data when event id is found, before discarding the event. This is a fail-safe setting in case something went wrong.timeout: (optional) timeout of unique event data stored in shm (in seconds), default 2. See theuniqueparameter of the post method.
The return value will be true, or nil and an error message.
This method can be called repeatedly to update the settings, except for the shm value which
cannot be changed after the initial configuration.
NOTE: the wait_interval is executed using the ngx.sleep function. In contexts where this
function is not available (eg. init_worker) it will execute a busy-wait to execute the delay.
syntax: is_already_configured = events.configured()
The events module runs as a singelton per workerprocess. The configured
function allows to check whether it is already up and running.
A check before starting any dependencies is recommended;
local events = require "resty.worker.events"
local initialization_of_my_module = function()
assert(events.configured(), "Please configure the 'lua-resty-worker-events' "..
"module before using my_module")
-- do initialization here
endsyntax: _M.events = events.event_list(sourcename, event1, event2, ...)
Utility function to generate event lists and prevent typos in
magic strings. Accessing a non-existing event on the returned table will result
in an 'unknown event error'.
The first parameter sourcename is a unique name that identifies the event
source, which will be available as field _source. All following parameters
are the named events generated by the event source.
Example usage;
local ev = require "resty.worker.events"
-- Event source example
local events = ev.event_list(
"my-module-event-source", -- available as _M.events._source
"started", -- available as _M.events.started
"event2" -- available as _M.events.event2
)
local raise_event = function(event, data)
return ev.post(events._source, event, data)
end
-- Post my own 'started' event
raise_event(events.started, nil) -- nil for clarity, no eventdata is passed
-- define my module table
local _M = {
events = events -- export events table
-- implementation goes here
}
return _M
-- Event client example;
local mymod = require("some_module") -- module with an `events` table
-- define a callback and use source modules events table
local my_callback = function(data, event, source, pid)
if event == mymod.events.started then -- 'started' is the event name
-- started event from the resty-worker-events module
elseif event == mymod.events.stoppping then -- 'stopping' is the event name
-- the above will throw an error because of the typo in `stoppping`
end
end
ev.register(my_callback, mymod.events._source)
syntax: success, err = events.poll()
Will poll for new events and handle them all (call the registered callbacks). The implementation is efficient, it will only check a single shared memory value and return immediately if no new events are available.
The return value will be "done" when it handled all events, "recursive" if it was
already in a polling-loop, or nil + error if something went wrong.
The "recursive" result simply means that an event-handler called poll again.
syntax: success, err = events.post(source, event, data, unique)
Will post a new event. source and event are both strings. data can be anything (including nil)
as long as it is (de)serializable by the cjson module.
If the unique parameter is provided then only one worker will execute the event,
the other workers will ignore it. Also any follow up events with the same unique
value will be ignored (for the timeout period specified to configure).
The process executing the event will not necessarily be the process posting the event.
The return value will be true when the event was successfully posted or
nil + error in case of failure.
Note: the worker process sending the event, will also receive the event! So if the eventsource will also act upon the event, it should not do so from the event posting code, but only when receiving it.
syntax: success, err = events.post_local(source, event, data)
The same as post except that the event will be local to the worker process,
it will not be broadcasted to other workers. With this method, the data element
will not be jsonified.
The return value will be true when the event was successfully posted or
nil + error in case of failure.
syntax: events.register(callback, source, event1, event2, ...)
Will register a callback function to receive events. If source and event are omitted, then the
callback will be executed on every event, if source is provided, then only events with a
matching source will be passed. If (one or more) event name is given, then only when
both source and event match the callback is invoked.
The callback should have the following signature;
syntax: callback = function(data, event, source, pid)
The parameters will be the same as the ones provided to post, except for the extra value
pid which will be the pid of the originating worker process, or nil if it was a local event
only. Any return value from callback will be discarded.
Note: data may be a reference type of data (eg. a Lua table type). The same value is passed
to all callbacks, so do not change the value in your handler, unless you know what you are doing!
The return value of register will be true, or it will throw an error if callback is not a
function value.
WARNING: event handlers must return quickly. If a handler takes more time than
the configured timeout value, events will be dropped!
Note: to receive the process own started event, the handler must be registered before
calling configure
syntax: events.register_weak(callback, source, event1, event2, ...)
This function is identical to register, with the exception that the module
will only hold weak references to the callback function.
syntax: events.unregister(callback, source, event1, event2, ...)
Will unregister the callback function and prevent it from receiving further events. The parameters work exactly the same as with register.
The return value will be true if it was removed, false if it was not in the handlers list, or
it will throw an error if callback is not a function value.
Nothing special is required, install like any other pure Lua module. Just make sure its location is in the module search path.
Please report bugs or submit patches by creating a ticket on the GitHub Issue Tracker,
Thijs Schreijer [email protected], Kong Inc.
This module is licensed under the Apache 2.0 license.
Copyright (C) 2016-2020, by Thijs Schreijer, Kong Inc.
All rights reserved.
- make sure changelog below is up-to-date
- update version number in the code
- create a new rockspec in
./rockspecs - commit with message
release x.x.x - tag the commit as
x.x.x - push commit and tags
- upload to luarocks
- chore: remove redundant type checking of unique_timeout.
- chore: add stacktrace to
posterrors for better debug information.
- fix: possible deadlock in the
initphase
- BREAKING: the
postfunction does not callpollanymore, making all events asynchronous. When an immediate treatment to an event is needed an explicit call topollmust be done. - BREAKING: the
post_localfunction does not immediately execute the event anymore, making all local events asynchronous. When an immediate treatment to an event is needed an explicit call topollmust be done. - fix: prevent spinning at 100% CPU when during a reload the event-shm is cleared
- fix: improved logging in case of failure to write to shm (add payload size for troubleshooting purposes)
- fix: do not log the payload anymore, since it might expose sensitive data through the logs
- change: updated
shm_retriesdefault to 999 - change: changed timer loop to a sleep-loop (performance)
- fix: when re-configuring make sure callbacks table is initialized
- feature: the polling loop now runs forever, sleeping for 0.5 seconds between runs, avoiding to create new timers on every step.
- BREAKING: the return values from
poll(and hence alsopostandpost_local) changed to be more lua-ish, to be truthy when all is well. - feature: new option
shm_retriesto fix "no memory" errors caused by memory fragmentation in the shm when posting events. - fix: fixed two typos in variable names (edge cases)
- fix: timeouts in init phases, by removing timeout setting, see issue #9
- change: add a stacktrace to handler errors
- fix: failing error handler if value was non-serializable, see issue #5
- fix: fix a test for the weak handlers
- OpenResty: http://openresty.org