GHSA-m43h-hfrq-x8wx SystemDS CPU exhaustion vulnerability

• Maven org.apache.systemds:systemds: < 2.2.2 → fixed in 2.2.2 • PyPI systemds: < 2.2.2 → fixed in 2.2.2

The Security Team noticed that the termination condition of the for loop in the readExternal method is a controllable variable, which, if tampered with, may lead to CPU exhaustion. As a fix, we added an upper bound and termination condition in the read and write logic. We classify it as a "low-priority but useful improvement". SystemDS is a distributed system and needs to serialize/deserialize data but in many code paths (e.g., on Spark broadcast/shuffle or writing to sequence files) the byte stream is anyway protected by additional CRC fingerprints. In this particular case though, the number of decoders is upper-bounded by twice the number of columns, which means an attacker would need to modify two entries in the byte stream in a consistent manner. By adding these checks robustness was strictly improved with almost zero overhead. These code changes are available in versions higher than 2.2.1.