A mostly seneca compatible, fast and stable microservice framework.
Seneca started degrading in performance and maintenance tended to be weak. We first tried maintaining it, but the code is mostly overcomplicated and hard to maintain.
This framework is being written for ourselves, first of all to replace seneca and second make it faster than it currently is and last but not least, make it stable. There were certain issues we faced and some missing parts which we partially added, like a circuit breaker. However this has again neither been really maintainable, nor stable.
Again this was created for us, no hate against seneca, they provided a great framework, we have used for years and we will speak out our support to them no matter what!
This codebase will be quite straightforward, plugins will be mostly skipped and we focus on the important part. Essential parts will be:
- discovery, via SWIM
- tcp transport (or quic if it ever gets released into node)
- alternative transports (the only plugin function we will provide for now)
- a faster pattern matching compatible algorithm
- a secondary
actE
function, which is more explicit in its behavior - fast method lookups
- fast duplicate call lookups
- load balancing
- circuit breaking
- native promises for act
We used actAsync
everywhere a promisified act
, if you used act
in your
project, this will break. You can monkey patch the act method of course very
easily like
const mc = new MicroWizard();
const act = mc.act;
mc.act = function (x, y, cb) {
if (typeof y === 'function') {
cb = y;
y = undefined;
}
if (typeof cb === 'function') {
act(x, y)
.catch(cb)
.then((x) => cb(null, x));
} else {
act(x, y).catch(console.error);
}
};
We have seneca widely in use, so we needed something compatible as replacement. Writing modules for one of the other frameworks wouldn't have fit the bill. So I decided to rewrite it and reuse parts of seneca were it makes sense.
I don't need to justify myself, but if you like to understand the reasoning:
The other frameworks including seneca are either missing features we need, or just have too much baggage in total. We need something rock solid that can scale to infinity. If possible it should be fast (seneca got slower and not well maintained, which is why we turned away from it). Fast means more throughput per service, which equals in less money spend and lower latencies. Both are very important to us.
The next thing is flexibility, how complicated it is to add things. In case of seneca, we consider the code to be actually unmaintainable, adding things is possible, but it wont live up to our standards, neither quality nor performance. This way we can add the necessary features. To name the most important Segmented Service discovery and fault detection by SWIM, load balancing with different algos + services themself can reject requests and get selected out temporarily by circuit breaking + tx retry (which is a loadbalancing strategy by itself). And general resilience against network errors with refeeding and deduplication.
Further we finally resolved issues we faced in some of our largest systems. Latency... . Due to time complexity in load balancing algorithms. Almost all non round robin algorithms tend to be quite costly. We usually have around 1-2k targets that go in a single load balancing algorithm. So this becomes quite heavy on high throughput and latency sensitive systems. We solve this with a linked list. This optimizes most of the things which tend to be issues when dealing with certain load balancing algorithms. The algorithm itself is round robin again, but with a few twists. If we want to add a weight to a certain machine, we assign an execution counter. The algorithm will then select the same target until the counter goes back to zero. If we encounter an overloaded or failing target, we circuit break the target and skip to the next (also ignoring any weights). So it is ultimately our services which tell requestors that they are over their threshold load, and not a central algorithm monitoring it. And last but not least the linked list allows very cheap adjustments of items. Either removing them, adding a new item, or updating their score. Removing has a time complexity of O(1), adding and moving an item O(log n). We could further optimize this with stored bins, but for our usecase this is more than enough, at least at the time of writing we have rather very short searches.
You can orient after the seneca docs to get an idea of how things are thought after. However here is the short guide.
You're using import
style (module/esm).
import MicroWizard from 'microwizard';
const mw = new MicroWizard();
// your code
You're using nodes cjs require
.
(async () => {
const { default: MicroWizard } = await import('microwizard');
const mw = new MicroWizard();
// your code
})();
There are two functions act
and actE
. actE
is faster but less flexible.
However most people will not need the flexibility of act
.
Let's explain how to add a function first
mw.add('service:test,command:math', (msg) => {
return { r: msg.a + msg.b };
});
And now consume it with act
const result = await mc.act(
'service:test',
{ command: 'math', a: 1, b: 2 },
{
// you can set sync to false, in this case the called function will not answer
// your call. In case your service is calling another module that is not in process
// the target service will also not send any answer back over the network.
//
// If you want an answer, you can ignore the meta param completely.
// sync = false;
sync: true // this is the default, you can omit this last param completely
}
);
console.log(result);
// returns { r: 3 }
This example quickly shows you all you need to know. The first parameter can receive a string or object, the same applies to the second. Everything on the root level of this two objects can be used to route to your target function. Read the seneca docs to understand this better.
!Important!. We only take the root level of the objects into consideration for finding a matching pattern.
Now finally what is different about actE
. This method works similar, but one important difference, only the first parameter is used by default for routing. Also the first paramis optimized to be fastest with jsonic like string (we do not support any quotes by the way which seneca does, the supported syntax is key:value[,key:value,...]
). So we recommend to use the string syntax for the first parameter.
The second parameter is used as our data object which contains the actual information forour target function.
The third parameter is not meta, but an options parameter. This options parameter
currently supports only one option which is mixin
. With mixin
you can specify an
array of keys, of root level elments of your data object that you want to add to the
pattern matching for the routing. This way we retain the flexibility of act, but with
more control and maximum performance. The tradeoff is a little bit of comfort, but you
will notice the few times you actually need to use mixin
is rare, which is exactly why
actE
exists in the first place, usually you statically adress what you want to call.
Inside the function you specified during add
you wont be able to access content on
request.msg directly, instead to access the content of the first param, you need to
access request.msg.msg
Everything else works like in act
. Here a few calls that show you the same call but in different ways:
const result = await mc.actE('service:test,command:math', { a: 1, b: 2 });
console.log(result);
// returns { r: 3 }
result = await mc.actE('service:test', { command: 'math' a: 1, b: 2 }, { mixin: ['command'] });
console.log(result);
// returns { r: 3 }
result = await mc.actE('service:test', { command: 'math' a: 1, b: 2 }, { mixin: ['command'] }, { sync: false });
console.log(result);
// returns immediately {} as the target service wont answer, if the service
// resides inside the same process it could theoretically answer though.
result = await mc.actE({ service: 'test' }, { command: 'math' a: 5, b: 2 }, { mixin: ['command'] });
console.log(result);
// returns { r: 7 }
// last but not least, the first param can be an object as well. However
// this is for dynamic purposes, and its not too slow as well, but the string based
// object definition will outperform, so use this when you need it and not by default
Look in the examples folder for a full project example. https://github.com/WizardTales/MicroWizard/tree/master/example/full
MicroWizard is mostly compatible with seneca, but there are a lot of differences.
So you can't expect any plugins to just work. The most important plugin is the
repl plugin. This can be made to work with @seneca/[email protected]
and the following code
'use strict';
const repl = require('@seneca/repl');
const jsonic = require('@jsonic/jsonic-next');
const Promise = require('bluebird');
const config = require('./config.js'); // optional
const dns = require('dns');
const os = require('os');
function dnsSeed(seneca, options, bases, next) {
dns.lookup(
config.baseName,
{
all: true,
family: 4
},
(err, addresses) => {
let bases = [];
if (err) {
throw new Error('dns lookup for base node failed');
}
if (Array.isArray(addresses)) {
bases = addresses.map((address) => {
return address.address;
});
} else {
bases.push(addresses);
}
next(bases);
}
);
}
(async () => {
const { default: MicroWizard } = await import('microwizard');
const seneca = new MicroWizard();
const initialSenecaConfig = {
auto: true,
discover: {
custom: {
active: true,
find: dnsSeed
}
}
};
const senecaConfig = {
...config.seneca,
...initialSenecaConfig
};
if (senecaConfig.bases && senecaConfig.bases.indexOf(',')) {
senecaConfig.bases = senecaConfig.bases.split(',');
}
seneca.use('mesh-ng', senecaConfig);
// we need to add this interface, it does absolutely nothing but makes
// the call repl expects work
seneca.init = (cb) => cb(() => {});
// we patch the original jsonic in, to have the full jsonic parser available
// in repl
seneca.util.Jsonic = jsonic;
// the monkey patch for act since the plugin expects of course the callback
// interface
const act = seneca.act;
seneca.act = function (x, y, cb) {
if (typeof y === 'function') {
cb = y;
y = undefined;
}
if (typeof cb === 'function') {
act
.call(seneca, x, y)
.catch(cb)
.then((x) => cb(null, x));
} else {
act.call(seneca, x, y).catch(console.error);
}
};
// and finally instead of calling seneca.use, we make the plugin work by
// defining its context and just calling it
repl.call(seneca, { ...repl.defaults, ...config.repl });
// seneca.use('@seneca/repl', config.repl);
})();
While this is not the most important part for us (of course this actually saves us quite some money), here are some benchmarks.
This is the test system:
Platform info:
==============
Linux 6.9.7-arch1-1 x64
Node.JS: 18.14.2
V8: 10.2.154.26-node.22
11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz × 16
Here the in process performance comparison of different acts.
✔ Moleculer* 2.096.783 rps
✔ Seneca* 25.743 rps
✔ MicroWizard* 529.986 rps
✔ MicroWizard2* 2.044.895 rps
✔ MicroWizard2Str* 3.422.563 rps
✔ MicroWizard2StrNoMix* 4.967.360 rps
✔ MicroWizardStr\* 675.521 rps
Moleculer* -57,79% (2.096.783 rps) (avg: 476ns)
Seneca* -99,48% (25.743 rps) (avg: 38μs)
MicroWizard* -89,33% (529.986 rps) (avg: 1μs)
MicroWizard2* -58,83% (2.044.895 rps) (avg: 489ns)
MicroWizard2Str* -31,1% (3.422.563 rps) (avg: 292ns)
MicroWizard2StrNoMix* 0% (4.967.360 rps) (avg: 201ns)
MicroWizardStr\* -86,4% (675.521 rps) (avg: 1μs)
---
Communicating over TCP, first an interpolated benchmark. The library used does not actually take care of concurrency. So we added it manually.
Seneca and Cote dropped their performance drastically down to a few rps (23) as soon as the concurrency raised over 2. Moleculer and our framework could sustain more than 10 in concurrency.
We were able to max out both Moleculer and our framework. However these are not final numbers. The way we created concurrency is by executing always a batch at the specified concurrency level. If we would have respawned a finished request immediately and upheld the concurrency, we could probably get even higher numbers on both of the top performers.
The methodology of this table is, that we multiplied the test result with the applied conccurency. We skip Cote in this table, since it failed already on a concurrency level of 2. So we show you now two tables for concurrency.
Concurrency level: 2
✔ Moleculer* 26.568 rps
✔ Seneca* 5.972 rps
✔ MicroWizardActE* 32.754 rps
Moleculer* -18,88% (26.568 rps) (avg: 37μs)
Seneca* -81,77% ( 5.972 rps) (avg: 167μs)
MicroWizardActE* 0% (32.754 rps) (avg: 30μs)
Here the high concurrency one.
Concurrency level: 1000
✔ Moleculer* 71.000 rps
✔ MicroWizardActE* 105.000 rps
Moleculer* -32,72% (71.000 rps) (avg: 6µs)
MicroWizardActE* 0% (105.000 rps) (avg: 9µs)
And to be complete, the last table. Running the test without setup concurrency.
Concurrency level: 1
-----------------------------------------------------------------------
Suite: Call remote actions
✔ Moleculer* 26.016 rps
✔ Cote* 34.258 rps
✔ Seneca* 5.610 rps
✔ MicroWizard1* 26.248 rps
✔ MicroWizard2* 26.844 rps
✔ MicroWizardActE* 30.667 rps
Moleculer* -24,06% (26.016 rps) (avg: 38μs)
Cote* 0% (34.258 rps) (avg: 29μs)
Seneca* -83,62% (5.610 rps) (avg: 178μs)
MicroWizard1* -23,38% (26.248 rps) (avg: 38μs)
MicroWizard2* -21,64% (26.844 rps) (avg: 37μs)
MicroWizardActE* -10,48% (30.667 rps) (avg: 32μs)
-----------------------------------------------------------------------
Here is the benchmark repo.