* Read replica update working set fixes
Read replica pull updates the session working set after pulling a
filtered set of branches from the tracking database. The original
(buggy) implementation updates the working set to the branch specified
at server-start time. The identity of that branch was fixed for the
duration of the server, so dolt_checkout would (appear to) have no
effect on the new branch's working set. What actually happened was more
pernicious: the working set was updated to the value of the
incorrect branch.
The fix no longer statically sets the active branch for a read replica
database. The active branch is pulled from the `*sql.Context`, so the
correct working set will be updated.
* [ga-format-pr] Run go/utils/repofmt/format_repo.sh and go/Godeps/update.sh
* add test
* todos
* delete extraeneous line
Co-authored-by: max-hoffman <max-hoffman@users.noreply.github.com>
* [no-release-notes] New enginetest harness
* touch ups
* [ga-format-pr] Run go/utils/repofmt/format_repo.sh and go/Godeps/update.sh
* missing dep
* fix new storage engine tests
* [ga-format-pr] Run go/utils/repofmt/format_repo.sh and go/Godeps/update.sh
* missing go.sum
* refactor setup package
* [ga-format-pr] Run go/utils/repofmt/format_repo.sh and go/Godeps/update.sh
* refactor setup package again
* [ga-format-pr] Run go/utils/repofmt/format_repo.sh and go/Godeps/update.sh
Co-authored-by: max-hoffman <max-hoffman@users.noreply.github.com>
Before this change, when an initial stage function returned a non-nil err, it
would return non-nil, resulting in closing its output channel and leaving the
errgroup responsible for canceling the egCtx.
The next transform stage would read `nil, false` on its inCh. That stage would
forward the `nil` down the transform pipeline. If the inCh selects all resolved
down the transform pipeline before any of the `<- ctx.Done()` selects, then the
pipeline would see the exact same behavior for the error case as it would see
for an `io.EOF` being reached in the first stage.
Instead, transform stages in the pipeline should abort without forwarding data
if they see a closed input channel. In turn, input stages should forward one
last `nil` sentinel value to their output channel when the see an `io.EOF`
before they close the channel.
Andy Arthur