Status
Current state: Accepted
Discussion thread: here (Not happening yet)
JIRA: here
Motivation
Instantiating a new client may result in a fatal failure if the bootstrap server cannot be resolved due to misconfiguration or transient network issues such as slow DNS. This is suboptimal because of the fact that it might take a long time for the address to become available at the DNS server, and users will need to continue to retry. Also, the ConfigException exception type does not accurately reflect the root cause of the problem, which makes it hard to handle this failure case. We think it is reasonable to allow users to have a grace period to retry if the address cannot be resolved immediately. Also, poisoning the clients during the construction can be obstructive; I think it is better to fail the client on its first attempt to connect to the network.
Proposed Changes
This KIP proposes moving bootstrapping logic from the constructor to the NetworkClient poll for two purposes,
1. not failing the client upon instantiation. In many cases, this behavior also kills the app, which might not be desirable.
2. piggybacking onto the client poll is a more natural way to retry.
We propose to add a new configuration option for timing out the bootstrapping process, a new exception type for handling bootstrap-related issues, and additional logging to aid in diagnosing bootstrapping failures.
Public Facing Changes
- Timeout Configuration: bootstrap.resolve.timeout.ms
- Exception: BootstrapResolutionException extends KafkaException
- Logging: log.WARN("Unable to bootstrap after {} ms.", elapsedMs)
Internal Changes
Client Constructor: The constructor will only parse the bootstrap configuration.
NetworkClient:
- Bootstrapping will now occur in the poll method before attempting to update the metadata. This includes resolving the addresses and bootstrapping the metadata.
- An error message will be logged in the event of a failed bootstrap process.
- If the timeout exceeds, a non-retriable BootstrapResolutionException will be thrown.
- Consumer, Producer, and Admin Clients: The bootstrap code will be changed.
New Configuration
bootstrap.resolve.timeout.ms
The proposed configuration specifies the maximum amount of time clients can spend trying to resolve for the bootstrap server address. If the resolution cannot be completed within this timeframe, a BootstrapResolutionException will be thrown.
Type: | long |
Default: | 120000 (i.e. 2 minutes) |
Valid Values: | 0 - LONG_MAX |
Importance: | high |
New Error
Name: BootstrapResolutionException extends KafkaException
Message: "Couldn't resolve server {} from {} as DNS resolution failed for {}"
Type: Non-retriable.
Compatibility, Deprecation, and Migration Plan
Compatibility
- Failed DNS resolution throws BootstraResolutionException: Users are expected to catch the error or the client will be poisioned
- Users who tried to catch ConfigException for DNS resolution error will no longer need this logic.
Deprecation
- There's no deprecation plan
Migration
- There's no migration plan
Case Study
In this section, I outlined how clients can react to bootstrap failures. In particular, I want to cover two common cases:
- Misconfiguration or non-transient issues with the network
- Transient network issues, e.g., slow DNS resolution.
KafkaConsumer
Case 1: Non-transient case
When the bootstrap timeout expires, the client will throw a BootstrapResolutionException.
Case 2: Transient Network Issue
consumer poll won't return any record until the client has been bootstrapped. If the issue cannot be resolved within the bootstrap timeout, a BootstrapConnectionException will be thrown.
KafkaProducer
Case 1: Non-transient case
The BootstrapResolutionException will be thrown in send() and partitionsFor() when the bootstrap timeout expires. If the max.block.ms elapsed before the timeout expires, a TimeoutException will be thrown instead.
Case 2: Transient Network Issue
The send() and partitionsFor() methods will be blocked on bootstrap until either the max.block.ms or the bootstrap timeout elapses.
AdminClient
Case 1: Non-transient case
The API call results will either timeout if the request times out first or be completed exceptionally with a BootstrapResolutionException.
Case 2: Transient Network Issue
The user won't be able to get the results back until the address is resolved. Meanwhile, the API calls can expire.
What should users do after the timeout expires?
The exception is meant to be fatal, so the user should check their network setup, configuration, or adjust the timeout.
The user can continue to retry, but this exception is meant to alert user to take action upon failing to bootstrap.
Test Plan
NetworkClient
Test DNS resolution upon its initial poll
Test if the right exception type is thrown
Existing clients (Consumer, Producer, AdminClient)
Test successful bootstrapping upon retrying
Rejected Alternatives
We've discussed many alternatives. Eventually, we asked ourselves what's the goal of this KIP, i.e., giving people a chance to retry on DNS resolution without poisoning the client. Which came down to two resolutions: 1. giving people a configurable timeout, and 2. adding a fatal error to alert the user.
Here are the rejected alternatives:
Maintain the current code behavior and add a retry loop with a timeout.
Pros: Same logic, less code change.
Cons: Do users want to be blocked on instantiating the client? I don't like this idea.
Throw DNS resolution upon failing but no retry
Pros: No additional config is needed
Cons: This is a behavioral change, and the application owner might need to rewrite the exception handling, i.e. catching the DNS failure logic.
- No retry. The network client will continue to retry until it is interrupted.
- Pros: No compatibility break. No additional exception handling logic, the network client will just log the error and continue to retry upon the next poll
- Cons: I think we should have some failure mechanism to notify users.
- Making BootstrapResolutionException retriable
- Pros: For the transient case, we might not even need a timeout, people are expected to retry on catching this exception
- Cons: Then we reply on alerting mechanism to alert users the issue. If it is indeed a configuration issue, then it is harder to discover
- Combine DNS resolution and connection into a single timeout
- Pros: Using a single timer to account for the connection time.
- Cons: Should we make connection retry fatal after the timeout? Maybe not.
- 5min default timeout
- We've decided to reduce it to 2min to stay coherent to the delivery.timeout.ms
- 5min can be too long