fix(storage): Offline Series Compaction

[benbjohnson]

This pull request adds the ability to offline compact existing series files into the influx_inspect buildtsi tool:

$ influx_inspect buildtsi -compact-series-file

It iterates over all series in each segment and rewrites all series that have not been tombstoned in the index to a new .tmp file next to the segment. Once all segments are converted, the temporary files overwrite the original segments.

The included benchmark shows that compacting 1M series takes only about 63ms:

% go test -bench=BenchmarkSeriesFile_Compaction -run=XXX -v ./tsdb
goos: darwin
goarch: amd64
pkg: github.com/influxdata/influxdb/tsdb
BenchmarkSeriesFile_Compaction-16    	      21	  62932487 ns/op
PASS
ok  	github.com/influxdata/influxdb/tsdb	14.363s

Describe your proposed changes here.

• CHANGELOG.md updated with a link to the PR (not the Issue)
• Well-formatted commit messages
• Rebased/mergeable
• Tests pass
• http/swagger.yml updated (if modified Go structs or API)
• Documentation updated or issue created (provide link to issue/pr)
• Signed CLA (if not already signed)

[e-dard]

I'm surprised that it's that fast. I seem to recall it being quite a bit slower when I prototyped it. Is the main improvement you made to add the segment compaction directly on the segment itself?

Couple of other questions:

1. what steps have been taken to test/verify this?
2. are you at all concerned about the size of the segments no longer being powers of two after compaction?

[benbjohnson]

@e-dard I'm testing on an 8-core i9 and it's running in parallel. I was surprised by the speed too though. I added a bunch of println() statements as I was benchmarking and it seems to be legit doing real work.

I'm OOO today and tomorrow but I'll add some additional test coverage. I did hit a bug with the use of FindOffsetByID() (which doesn't check tombstones) vs IsDeleted() (which does check tombstones).

I'm not concerned about the sizing being off. That was originally added so that we could create a file with a fixed size for Windows.

[benbjohnson]

Added test coverage for the series compaction. I'm getting a CI error from Semantic Pull Request but I can't figure out why.

[dgnorton]

@benbjohnson the PR title also has to be in the same format

[e-dard]

@8none1 I think we should get some real-world testing of this under our belts before 1.8 ships. Perhaps by taking a real workload and verifying the tool does the right thing.

[e-dard]

LGTM 👍