We touched on this a bit in previous discussions, but I wanted to draw out the approach to config specifically as an item of discussion.
The new producer and consumer use a similar key-value config approach as the existing scala clients but have different implementation code to help define these configs. The plan is to use the same approach on the server, once the new clients are complete; so if we agree on this approach it will be the new default across the board.
Let me split this into two parts. First I will try to motivate the use of key-value pairs as a configuration api. Then let me discuss the mechanics of specifying and parsing these. If we agree on the public api then the public api then the implementation details are interesting as this will be shared across producer, consumer, and broker and potentially some tools; but if we disagree about the api then there is no point in discussing the implementation.
Let me explain the rationale for this. In a sense a key-value map of configs is the worst possible API to the programmer using the clients. Let me contrast the pros and cons versus a POJO and motivate why I think it is still superior overall.
Pro: An application can externalize the configuration of its kafka clients into its own configuration. Whatever config management system the client application is using will likely support key-value pairs, so the client should be able to directly pull whatever configurations are present and use them in its client. This means that any configuration the client supports can be added to any application at runtime. With the pojo approach the client application has to expose each pojo getter as some config parameter. The result of many applications doing this is that the config is different for each and it is very hard to have a standard client config shared across. Moving config into config files allows the usual tooling (version control, review, audit, config deployments separate from code pushes, etc.).
Pro: Backwards and forwards compatibility. Provided we stick to our java api many internals can evolve and expose new configs. The application can support both the new and old client by just specifying a config that will be unused in the older version (and of course the reverse--we can remove obsolete configs).
Pro: We can use a similar mechanism for both the client and the server. Since most people run the server as a stand-alone process it needs a config file.
Pro: Systems like Samza that need to ship configs across the network can easily do so as configs have a natural serialized form. This can be done with pojos using java serialization but it is ugly and has bizare failure cases.
Con: The IDE gives nice auto-completion for pojos.
Con: There are some advantages to javadoc as a documentation mechanism for java people.
Basically to me this is about operability versus niceness of api and I think operability is more important.
Let me now give some details of the config support classes in kafka.common.config and how they are intended to be used.
The goal of this code is the following: 1. Make specifying configs, their expected type (string, numbers, lists, etc) simple and declarative 2. Allow for validating simple checks (numeric range checks, etc) 3. Make the config "self-documenting". I.e. we should be able to write code that generates the configuration documentation off the config def. 4. Specify default values. 5. Track which configs actually get used. 6. Make it easy to get config values.
There are two classes there: ConfigDef and AbstractConfig. ConfigDef defines the specification of the accepted configurations and AbstractConfig is a helper class for implementing the configuration class. The difference is kind of like the difference between a "class" and an "object": ConfigDef is for specifying the configurations that are accepted, AbstractConfig is the base class for an instance of these configs.
(Ignore the static config names in there for now...I'm not actually sure that is the best approach).
So the way this works is that the config specification is defined as:
config = new ConfigDef().define("bootstrap.brokers", Type.LIST, "documentation")
.define("metadata.timeout.ms", Type.LONG, 60 * 1000, atLeast(0), "documentation") .define("max.partition.size", Type.INT, 16384, atLeast(0), "documentation") This is used in a ProducerConfig class which extends AbstractConfig to get access to some helper methods as well as the logic for tracking which configs get accessed.
Currently I have included static String variables for each of the config names in that class. However I actually think that is not very helpful as the javadoc for them doesn't give the constant value and requires duplicating the documentation. To understand this point look at the javadoc and note that the doc on the string is not the same as what we define in the ConfigDef. We could just have the javadoc for the config string be the source of truth but it is actually pretty inconvient for that as it doesn't show you the value of the constant, just the variable name (unless you discover how to unhide it). That is fine for the clients, but for the server would be very weird especially for non-java people. We could attempt to duplicate documentation between the javadoc and the ConfigDef but given our struggle to get well-documented config in a single place this seems unwise.
So I recommend we have a single source for documentation of these and that that source be the website documentation on configuration that covers clients and server and that that be generated off the config defs. The javadoc on KafkaProducer will link to this table so it should be quite convenient to discover. This makes things a little more typo pro
Overall, +1 on sticking with key-values for configs.
Optionally, both the above cons can be addressed (to some degree) by wrapper config POJOs that read in the config. i.e., the client will provide a KV config, but then we (internally) would load that into a specific config POJO that will be helpful for auto-completion and javadocs and convenience for our internal implementation (as opposed to using getLong/getString, etc. which could cause runtime exceptions if done incorrectly). The javadoc in the pojo would need a @value link to the original config key string if it is to show up in the generated javadoc.
Figuring out a way to un-hide it would be preferable to having to keep the website as the single source of documentation (even if it is generated from the javadoc) and make the javadoc link to it. I tried, but was unsuccessful so unless someone knows how to do that the above approach is the next-best alternative.
I like the helper function in all except in parseType: is it better to be restrict about types, i.e. now allowing "true" if the type is really Boolean? On Wed, Feb 5, 2014 at 5:06 PM, Joel Koshy <[EMAIL PROTECTED]> wrote:
In spirit I agree with you, but it is really annoying if I can't say props.put("config.a", 45) props.put("config.b", true)
This is especially true because config will actually come from elsewhere and may have the appropriate type so in reality you end up with props.put("config.a", Integer.toString(theValue)) and the error is usually pretty confusing because it is like "Invalid value true, not a boolean!"
Since the type is specified I can't think of any corner case with confusing behavior.
-Jay On Wed, Feb 5, 2014 at 5:30 PM, Guozhang Wang <[EMAIL PROTECTED]> wrote:
Ah, I actually don't think the internal usage is a problem for *us*. We just use config in one place, whereas it gets set in 1000s of apps, so I am implicitly optimizing for the application interface. I agree that we can add getters and setters on the ProducerConfig if we like. Basically I was just concerned about the user, the nice thing about a pojo is that when you type config.set The ide pops up a list of configs with documentation right there, which is just very convenient.
-Jay On Wed, Feb 5, 2014 at 5:06 PM, Joel Koshy <[EMAIL PROTECTED]> wrote:
You raise a couple points, let me give the rationale and see if you are convinced.
1. Why not reuse commons config? This is exactly the right question for application programming. We could save ourselves a few hundred lines of code we would maintain by just adding a library dependency, which is almost always the right decision if you are building an application as the application is itself the end result. But the issue here is that these clients will be embedded in thousands of applications which will in turn have other dependencies. So as a result any common libraries we use will also be used by others and this inevitably leads to dependency/incompatibility hell. As a result I really think for this kind of helper code we should just do it ourselves. I think of this as kind of our sacrifice for the users. :-)
Meanwhile since we are decoupling the client and the server and since the server should always be a stand alone thing we should be able to be much more liberal in our dependencies there.
2. The next question is why not use a commons-config-like approach where the config object is a map/properties thing and we wrap it in a pojo that provides getters that parse values. This is the approach we took in the existing client and server code. The problem with this is that it is impossible to determine the set of configs programmatically. The goal was to have our config documentation automatically generated off the code (including the type, default value, documentation, etc). This also has the nice side effect that all the configs are validated up front, you don't have to wait until someone calls the getter to check validity.
3. This approach doesn't prohibit custom validation. All validation is done with a Validator, so you can plug in any validation you like. This can be a bit ugly for one-off code (annonymous inner class, yuck...but will get better with java 8 and in scala is already good). But we can also just do validation in line in the wiring class. I see the "responsibility" of KafkaProducer or KafkaServer classes as turning configs into a constructed assembled hierachy of objects so checking stuff there is totally legit.
4. If the pojo wrapper you describe is just an internal convenience for our code (as Joel described), then I have no objection to that. But the objection to a user facing pojo was what I described before...
-Jay On Thu, Feb 6, 2014 at 6:27 AM, David Arthur <[EMAIL PROTECTED]> wrote:
I made a typo in my previous email, I meant to say "not allowing", i.e. if a config' type is Boolean, do not accept "true" as the value. On Thu, Feb 6, 2014 at 8:30 AM, Jay Kreps <[EMAIL PROTECTED]> wrote:
You have to accept the boolean true to support convenient programmatic configuration (the example I showed). You have to support taking a string to allow injecting config from a config file or external application config system. E.g. in the case where you supply a Properties instance if that instance was read from a .properties file the value will always be a string.
-Jay On Thu, Feb 6, 2014 at 9:51 AM, Guozhang Wang <[EMAIL PROTECTED]> wrote:
The Configurable interface was meant to use with plugins. Originally we had two of these in the producer, Serializer and Partitioner but we removed them so now it is unused. However I think we will likely have several in the consumer so I will leave it there for now. The idea was for interfaces where the user provides an implementation you often need some configuration in your implementation (e.g. a range partitioner may need to fetch its list of break points, a string serializer may need to take a configurable charset, etc). That interface is a way to pass in the configuration when instantiating the object.
In the existing scala code we have the same basic idea but there is no configure method, instead we just look for a constructor that takes properties. The problem with this is that it isn't very discoverable so no one knows how to use it.
That is a good point that this approach does assume static configuration. I actually prefer this because we follow a dependency injection style of config usage which allows you to take values and hold them in final/val variables. Getting away from this to allow any config to change dynamically is pretty hard. I think dynamic config tends to have a lot of corner cases and synchronization to handle correctly. Some things are very difficult to dynamically configure and others are impossible (like socket buffers)). My preference would just be to work on making startup/shutdown quick.
-Jay On Thu, Feb 6, 2014 at 11:52 AM, David Arthur <[EMAIL PROTECTED]> wrote:
+1 on the new config. Just one comment. Currently, when initiating a config (e.g. ProducerConfig), we log those overridden property values and unused property keys (likely due to mis-spelling). This has been very useful for config verification. It would be good to add similar support in the new config.
Jun On Tue, Feb 4, 2014 at 9:34 AM, Jay Kreps <[EMAIL PROTECTED]> wrote:
I think that is reasonable but would object to having it be debug logging? I think logging out a bunch of noise during normal operation in a client library is pretty ugly. Also, is there value in exposing the final configs programmatically?
On Sun, Feb 9, 2014 at 9:23 PM, Jun Rao <[EMAIL PROTECTED]> wrote:
I actually prefer to see those at INFO level. The reason is that the config system in an application can be complex. Some configs can be overridden in different layers and it may not be easy to determine what the final binding value is. The logging in Kafka will serve as the source of truth.
For reference, ZK client logs all overridden values during initialization. It's a one time thing during starting up, so shouldn't add much noise. It's very useful for debugging subtle config issues.
Exposing final configs programmatically is potentially useful. If we don't want to log overridden values out of box, an app can achieve the same thing using the programming api. The only missing thing is that we won't know those unused property keys, which is probably less important than seeing the overridden values.
Jun On Mon, Feb 10, 2014 at 10:15 AM, Jay Kreps <[EMAIL PROTECTED]> wrote:
I actually think this is useful for non-LinkedIn users as well. The following is the tradeoff that I see.
Most users probably won't care about seeing an extra 10-20 lines of INFO level logging when starting up a client. However, it's very easy for users to (1) mis-spell a config name (there was an issue in the mailing list just a few days ago when a user mistyped "advertised.host" to "advertise.host") or (2) inadvertently override a config value through the config systems. In both cases, the INFO level logging will make it much easier for the user to realize the human mistake. So, I think this is a case where the benefit outweighs the disadvantage.
Jun On Mon, Feb 10, 2014 at 4:13 PM, Jay Kreps <[EMAIL PROTECTED]> wrote:
Apache Lucene, Apache Solr and all other Apache Software Foundation project and their respective logos are trademarks of the Apache Software Foundation.
Elasticsearch, Kibana, Logstash, and Beats are trademarks of Elasticsearch BV, registered in the U.S. and in other countries. This site and Sematext Group is in no way affiliated with Elasticsearch BV.
Service operated by Sematext