Crates.io | streambed-logged |
lib.rs | streambed-logged |
version | 0.11.1 |
source | src |
created_at | 2023-10-18 04:08:35.502083 |
updated_at | 2024-09-13 03:32:41.207796 |
description | Logged is a small library that implements a file-system-based commit log |
homepage | |
repository | https://github.com/streambed/streambed-rs.git |
max_upload_size | |
id | 1006357 |
size | 106,742 |
Logged is a small library that implements a file-system-based commit log function for autonomous systems that often live at the edge of a wider network.
Logged implements the Streambed commit log API which, in turn, models itself on the Kafka API.
Nothing beats code for a quick introduction! Here is an example of establishing the commit log, producing a record and then consuming it. Please refer to the various tests for more complete examples.
let cl = FileLog::new(logged_dir);
let topic = Topic::from("my-topic"_;
cl.produce(ProducerRecord {
topic: topic.clone(),
headers: None,
timestamp: None,
key: 0,
value: b"some-value".to_vec(),
partition: 0,
})
.await
.unwrap();
let subscriptions = vec![Subscription {
topic: topic.clone(),
}];
let mut records = cl.scoped_subscribe("some-consumer", None, subscriptions, None);
assert_eq!(
records.next().await,
Some(ConsumerRecord {
topic,
headers: None,
timestamp: None,
key: 0,
value: b"some-value".to_vec(),
partition: 0,
offset: 0
})
);
The primary functional use-cases of logged are:
The primary operational use-cases of logged are:
Logged has no notion of what a network is. Any replication of the data managed by logged is to be handled outside of it.
Logged is optimized for small memory usage over speed and also recognizes that storage is limited.
Logged deliberately constrains the appending of a commit log to one process. This is known as the single writer principle, and greatly simplifies the design of logged.
Readers can exist in processes other than the one that writes to a commit log topic. There is no broker process involved as both readers and writers read the same files on the file system. No locking is required for writing due to the single writer principle.
Encryption and compression are an application concern, with encryption being encouraged. Compression may be less important, depending on the volume of data that an application needs to retain.
Authentication is bypassed given the assumption of encryption. If a reader is unable to decrypt a message then this is seen to be as good as not gaining access in the first place. Avoiding authentication yields an additional side-effect of keeping logged free of complexity by not having to manage identity.
Logged provides the tools to the application so that an application may perform compaction. Our experience with applying commit logs has shown that data retention strategies are an application concern often subject to the contents of the data.
Logged is modelled with the same concepts as Apache Kafka including topics, partitions, keys, records, offsets, timestamps and headers. Services using logged may therefore lend themselves to portability toward Kafka and others.
A CRC32C checksum is used to ensure data integrity against the harsh reality of a host being powered off abruptly. Any errors detected, including incomplete writes or compactions, will be automatically recovered.
Logged is implemented as a library avoiding the need for a broker process. This is similar to Chronicle Queue on the JVM.
The file system is used to store records in relation to a topic and a single partition. Tokio is used for file read/write operations so that any stalled operations permit other tasks to continue running. Postcard is used for serialization as it is able to conveniently represent in-memory structures and is optimized for resource-constrained targets.
When seeking an offset into a topic, logged will read through records sequentially as there is no indexing. This simple approach relies on the ability for processors to scan memory fast (which they do), and also for the application-level compaction to be effective. By "effective" we mean that scanning can avoid traversing thousands of records.
The compaction of topics is an application concern and as it is able to consider the contents of a record. Logged provides functions to atomically "split" an existing topic log at its head and yield a new topic to compact to. The application-based compactor can then consume the existing topic, retain the records it needs in memory. Having consumed the topic, the application-based compactor will append the retained records to the topic to compact to. Once appended, logged can be instructed to replace the compacted log for the old one. Meanwhile, any new appends post splitting the log will be written to another new log for the topic. As a result, there can be two files representing the topic, and sometimes three during a compaction activity. Logged takes care of interpreting the file system and various files used when an application operates with a topic.
When you have Kafka.