Rotate compaction output at user key boundary

so that keys with same user key will not spread over multiple tables in
same level, level 1 or above.

Level compaction, except for level-0, don't pick overlapping tables. So
if we produce overlapping tables in user key level, it is possible that
file with newer version user key got picked and compacted to next level.
Now, the invariant that Version::Get depends on is broken, Version::Get
is broken.

The vulnerability of this approach is that a malicious client can retain
a snapshhot and do large number of writes to same key, which can cause
huge sorted table generated. LevelDB is a library that embedded in other
application. I think it should be the application's responsibility to
keep snapshhot from malicious clients.

Picking overlapping tables in all level compactions is another approach
to solve this problem. I think this approach may violate the algorithm
this implementation implies.

db/db_impl.cc

@@ -926,16 +926,10 @@ Status DBImpl::DoCompactionWork(CompactionState* compact) {
  }
 
  Slice key = input->key();
- if (compact->compaction->ShouldStopBefore(key) &&
- compact->builder != NULL) {
- status = FinishCompactionOutputFile(compact, input);
- if (!status.ok()) {
- break;
- }
- }
 
  // Handle key/value, add to state, etc.
  bool drop = false;
+ bool first_occurrence = false;
  if (!ParseInternalKey(key, &ikey)) {
  // Do not hide error keys
  current_user_key.clear();
@@ -947,6 +941,8 @@ Status DBImpl::DoCompactionWork(CompactionState* compact) {
  Slice(current_user_key)) != 0) {
  // First occurrence of this user key
  current_user_key.assign(ikey.user_key.data(), ikey.user_key.size());
+ // Don't rotate at this key if last key is an error key.
+ first_occurrence = last_sequence_for_key != kMaxSequenceNumber;
  has_current_user_key = true;
  last_sequence_for_key = kMaxSequenceNumber;
  }
@@ -980,6 +976,19 @@ Status DBImpl::DoCompactionWork(CompactionState* compact) {
 #endif
 
  if (!drop) {
+ // Close output file if it is overlapping too much grandparents
+ // or big enough. We never put keys with same user key to
+ // different files in same level in level compaction.
+ if (compact->builder != NULL && first_occurrence &&
+ (compact->compaction->ShouldStopBefore(key) ||
+ compact->builder->FileSize() >=
+ compact->compaction->MaxOutputFileSize())) {
+ status = FinishCompactionOutputFile(compact, input);
+ if (!status.ok()) {
+ break;
+ }
+ }
+
  // Open output file if necessary
  if (compact->builder == NULL) {
  status = OpenCompactionOutputFile(compact);
@@ -992,15 +1001,6 @@ Status DBImpl::DoCompactionWork(CompactionState* compact) {
  }
  compact->current_output()->largest.DecodeFrom(key);
  compact->builder->Add(key, input->value());
-
- // Close output file if it is big enough
- if (compact->builder->FileSize() >=
- compact->compaction->MaxOutputFileSize()) {
- status = FinishCompactionOutputFile(compact, input);
- if (!status.ok()) {
- break;
- }
- }
  }
 
  input->Next();