s/segment: fixed tracking offsets of empty segments

When a segment is rolled due to the `segment.ms` property its base
offset is set to the committed offset of last segment plus one. In a
segment constructor all segment offset tracker offsets were set to the
base offset passed as the constructor argument. This way an empty
segment had exactly the same set of offsets as the segment containing
single batch containing one record and base offset equal to the segment
base offset.

Incorrect offset accounting in empty segments lead to the situation in
which eviction driven `log::truncate_prefix` was called with an
incorrect offset, leading to the situation in which a log wasn't
truncated at the segment boundary. This in the end resulted in the
disconnection between the offset translator truncation point and first
log segment start offset.

For example:

(we represent segment as `[base_offset,end_offset]`)

In the case of log with segments:

```
[0,10][11,15]
```

After segment ms a new segment is created like this:

```
[0,10][11,15][16,16]
```

When eviction point is established the last offset is checked in this
case if all the segments are to be evicted it will be equal to 16.
`16` is a last offset that is going to be evicted (last included in raft
snapshot). Hence the `log::truncate_prefix` is called with offset `17`.

If some batches are appended to the rolled segment then the log will
contain f.e.

```
 [0,10][11,15][16,25]
```

Now if the log is prefix truncated at offset `17` it starts offset is
updated to `17` but the underlying segment is kept.

```
[16,25]
```

Now when the reader start reading the log it will request to start from
the log start offset which is `17` but it will have to skip over the
batch starting at `16`. If a batch at `16` has more than one record it
will still be returned to the reader and it will have to translate its
base offset to kafka offset space. This is impossible as the
offset_translator was already truncated to `16`. i.e. there is no
information in the translator to correctly translate the offset.

A fix is setting the empty segment dirty, committed and stable offset to
its base_offset minus one. This way all the semantics of log operation
holds as the offset returned is equal to previous segment last offset
which would be the case if there would be no empty segment at the head
of the log.

Fixes: redpanda-data#11403

Signed-off-by: Michal Maslanka <[email protected]>

src/v/storage/segment.cc

@@ -429,7 +429,7 @@ ss::future<> segment::do_truncate(
 
 ss::future<bool> segment::materialize_index() {
  vassert(
- _tracker.base_offset == _tracker.dirty_offset,
+ _tracker.base_offset == model::next_offset(_tracker.dirty_offset),
  "Materializing the index must happen before tracking any data. {}",
  *this);
  return _idx.materialize_index().then([this](bool yn) {

src/v/storage/segment.h

@@ -45,9 +45,9 @@ class segment {
  offset_tracker(model::term_id t, model::offset base)
  : term(t)
  , base_offset(base)
- , committed_offset(base)
- , dirty_offset(base)
- , stable_offset(base) {}
+ , committed_offset(model::prev_offset(base))
+ , dirty_offset(model::prev_offset(base))
+ , stable_offset(model::prev_offset(base)) {}
  model::term_id term;
  model::offset base_offset;