bug(source): schema registry header of protobuf is not skipped properly

[xiangjinwu]

This is skipping i bytes rather than skipping i variable-length zig-zag encoded integers. For i within 1..=63, the encoding outputs a single byte so this happens to work.

prefixed by the length of the array (which is also variable length, Zigzag encoded)

i itself is also encoded so the decoded length shall be i / 2.

Originally posted by @xiangjinwu in #14057 (comment)

Got it, so the correct implementation is

        Some(i) => Ok(&remained[(1 + *i / 2) as usize..]),

Originally posted by @Rossil2012 in #14057 (comment)

Impact

In the common case where a message is defined at top level (i.e. not nested within another message's definition), and the message index is <= 63, the current implement happens to work correctly because i == 2 implies i == 1 + i/2.

Example: UserWithNewType in the PR is at top level with index 2, so its indexes array [2] would be encoded as 02 04 (02 for the length 1 and 04 for the element 2). Skipping 02 (first byte read as i) bytes happens to skip the encoded sequence exactly.

[Rossil2012]

Note here:
As nesting depth of message and no. of messages can be larger than 2^6 - 1 (in varint encoding, only 7b out of 1B is for zig-zag encoding, so 63 is the max number that can be encoded in 1 byte), we need to check more bytes if the highest bit is 1.
The correct implementation may be:
if the first byte is not 0, decode one int as i with varint & zig-zag, and then decode i ints with varint & zig-zag. The remaining part is the actual payload.