The detailed post is here: https://flink.apache.org/2026/06/26/announcing-native-s3-fs/
If you use Flink you should check out the native S3 connector in Flink 2.3. There is a significant throughput improvement for state and checkpoint management if you use an object store to save it. The published benchmark covers a 20 GB RocksDB state with full non-incremental checkpoints every 60 seconds and compares the native plugin against the Presto plugin.
a) Average throughput improved from about 92 MB/s to about 200 MB/s.
b) Average checkpoint duration improved from 90.1 seconds to 48.8 seconds and P99 improved from 165.3 seconds to 76.7 seconds.
c) The job completed 78 checkpoints in the same window versus 40 with the Presto plugin.
d) The gain as you would expect is higher at smaller state and it reaches 4.58x in the 0 to 2 GB range and it stays at 2x or higher as the state grows.
The connector is built directly on AWS SDK v2 and that is worth paying attention to because SDK v1 reached end of support on 31 December 2025 and both existing plugins depend on it. The other technical benefits are these.
a) It is much lighter because it carries no Hadoop dependencies. The JAR is around 13 MB against 30 MB for the Hadoop plugin and 93 MB for the Presto plugin and that also removes the CVE triage work on everything Hadoop pulls in with it.
b) Reads and writes use Netty NIO so the connection is multiplexed instead of one thread per request. Bulk state restore runs as batched concurrent transfers inside the JVM with no external tooling needed (s5cmd).
c) It is the first plugin that gives you exactly-once file sinks and the faster checkpoint path in one place because until now you had to pick one or the other.