Have bank txn test be run like Jepsen, so different servers have different clients, each client running txns serially.

Make Dgraph server gracefully stop various goroutines in groups.go and draft.go, and cleanly shut down after completing the pending tasks.

Introduce a new /assignIds HTTP endpoint in Zero, so users can allocate UIDs to nodes externally to Dgraph.

* Correctly handle list of type geo in json (#2482)

* Add handleGeoType function for handling a geotype or list of geotype (#2482)

It is currently possible to start zero with --replicas=4 or --replicas=0
Even number means it's impossible to reach quorum.
No replicas doesn't make much sense at all.

Edits for wiki mutations page about JSON

update the wiki

Mainly edits on top of Michel's changes.

- Changed few notice notes.
- Add topic "Creating a list with JSON and interacting with"
- Add topic "JSON Syntax using Raw HTTP or Ratel UI".
- Add topic "Using JSON operations via cURL"

This makes LiveLoad faster by making concurrent requests optional.

The Blank Node should be the same, for connecting <loc> and <name>.

- Use exponential backoffs with propose txns, so if the system is flooded, we don't make it worse.
- Do not block if we find pending txns. Return err, and abort them in a goroutine.
- Do not reach a deadlock trying to remove posting lists from LRU cache. Do a best-effort removal.
- Decrease the number of times SyncIfDirty happens.

Remove trigram limit.

- Fix up proposeTxn, so we are updating the CommitTs as per the status in Zero oracle.
- Fix #2474 by dealing with a nil resp.Context in server.go
- Fix calculateSnapshot. It is not a given that every StartTs in Mutation proposal would have a corresponding commit/abort decision. So, use Oracle's MinPendingStartTs to decide how to calculate snapshot.
- Remove o.aborts in Zero. The work can be done by o.commits map.

Fix an issue with calculating snapshots, so we only consider mutations with edges, and not schema updates. Schema updates do not have a corresponding txn commit.

Txn:

- Commits have been separated into CommitToMemory and CommitToDisk. This logic is a lot simpler and clearer than before. In particular, CommitToDisk does not try to generate any state. It just serializes the posting list stored at startTs and pushes that to Badger.
- Multiple txns doing CommitToDisk can now be batched up into fewer transactions and use callbacks to execute quickly. Added a helper struct for that.
- Remove the concept of activeTxns in posting/list.go. It's a repetition.
- Simplify txn.deltas, and removed as much as I could.
- Removed abortMutation methods from txn (mvcc.go) and posting/list.go

Zero:

- TryAbort returns OracleDelta, which can be directly proposed in the group.
- abortOldTransactions uses this to ensure that in case we missed a transaction update, the result of the abort can then be proposed to the group, so group snapshot doesn't get stuck.

Fix up both the way we create a snapshot and the way we stream it. Make the logic robust and add tests to ensure that the calculation of snapshot is right.

Streaming no longer needs to worry about `min_ts`. We no longer send a snapshot upfront. Raft would request it as it desires, and the follower sends the snapshot details it needs from the leader. The leader in turn checks if its snapshot matches with what the follower wants. This is a better logic, but not entirely foolproof (comment in predicate.go explains).

Bring in Badger changes, which add a `DropAll` method, which is called before applying the snapshot.

Instead of discarding all versions anytime, we now only discard versions which are below the timestamp used by snapshot. Thus, we keep all versions of keys after the last snapshot.