The added test describes how the new id timestamps fix the race condition.
These timestamps could be any unique opaque values, and the protocol is
extended to exchange them along with node ids.

Internally, nodes also reuse timestamps as a marker to identify the first
NotifyNodeInformation packets from the master: since this packet is a complete
list of nodes in the cluster, any other node in the node manager has left the
cluster definitely and is removed.

The secondary masters didn't receive update about master nodes.
It's also useless to send them information about non-master nodes.

When Client (including backup master) and admin nodes are identified,
the primary master now sends them automatically all nodes with
NotifyNodeInformation, as with storage nodes.

- check address conflicts
- on invalid values, reject peer instead of dying

Listing connected/connecting nodes with a UUID is used:
- in one place by storage nodes: here, it does not matter if we skip nodes that
  aren't really identified
- in many places by the master, only for server connections, in which case we
  have equivalence with real identification

So in practice, NodeManager is only simplified to reuse the 'identified'
property of nodes.

Therefore, a client node in the node manager is always RUNNING.

Although the change applies to any node with a temporary ids (all but storage),
only clients don't have addresses and are therefore not recognizable.

After a client is disconnected from the master and before reconnecting, another
client may join the cluster and "steals" the id of the first client. This issue
leads to stuck clients, failing in loop with exceptions like the following one:

    ERROR ZODB.Connection Couldn't load state for 0x0251
    Traceback (most recent call last):
      File "ZODB/Connection.py", line 860, in setstate
        self._setstate(obj)
      File "ZODB/Connection.py", line 901, in _setstate
        p, serial = self._storage.load(obj._p_oid, '')
      File "neo/client/Storage.py", line 82, in load
        return self.app.load(oid)[:2]
      File "neo/client/app.py", line 353, in load
        data, tid, next_tid, _ = self._loadFromStorage(oid, tid, before_tid)
      File "neo/client/app.py", line 373, in _loadFromStorage
        for node, conn in self.cp.iterateForObject(oid, readable=True):
      File "neo/client/pool.py", line 91, in iterateForObject
        pt = self.app.pt
      File "neo/client/app.py", line 145, in __getattr__
        self._getMasterConnection()
      File "neo/client/app.py", line 214, in _getMasterConnection
        result = self.master_conn = self._connectToPrimaryNode()
      File "neo/client/app.py", line 246, in _connectToPrimaryNode
        handler=handler)
      File "neo/lib/threaded_app.py", line 154, in _ask
        _handlePacket(qconn, qpacket, kw, handler)
      File "neo/lib/threaded_app.py", line 135, in _handlePacket
        handler.dispatch(conn, packet, kw)
      File "neo/lib/handler.py", line 66, in dispatch
        method(conn, *args, **kw)
      File "neo/lib/handler.py", line 188, in error
        getattr(self, Errors[code])(conn, message)
      File "neo/client/handlers/__init__.py", line 23, in protocolError
        raise StorageError("protocol error: %s" % message)
    StorageError: protocol error: already connected

`ClientCache._oid_dict` shall not have empty values. For a given oid, when the
last item is removed from the cache, the oid must be removed as well to free
memory. In some cases, this was not done.

A consequence of this bug is the following exception:

    ERROR ZODB.Connection Couldn't load state for 0x02d1e1e4
    Traceback (most recent call last):
      File "ZODB/Connection.py", line 860, in setstate
        self._setstate(obj)
      File "ZODB/Connection.py", line 901, in _setstate
        p, serial = self._storage.load(obj._p_oid, '')
      File "neo/client/Storage.py", line 82, in load
        return self.app.load(oid)[:2]
      File "neo/client/app.py", line 358, in load
        self._cache.store(oid, data, tid, next_tid)
      File "neo/client/cache.py", line 228, in store
        prev = item_list[-1]
    IndexError: list index out of range

A backup cluster for tids <= backup_tid has all data to provide regular
read-only ZODB service. Having regular ZODB access to the data can be
handy e.g. for externally verifying data for consistency between
main and backup clusters. Peeking around without disturbing main
cluster might be also useful sometimes.

In this patch:

- master & storage nodes are taught:

    * to instantiate read-only or regular client service handler depending on cluster state:
      RUNNING   -> regular
      BACKINGUP -> read-only

    * in read-only client handler:
      + to reject write-related operations
      + to provide read operations but adjust semantic as last_tid in the database
        would be = backup_tid

- new READ_ONLY_ACCESS protocol error code is introduced so that client can
  raise POSException.ReadOnlyError upon receiving it.

I have not implemented back-channel for invalidations in read-only mode (yet ?).
This way once a client connects to cluster in backup state, it won't see
new data fetched by backup cluster from upstream after client connected.

The reasons invalidations are not implemented is that for now (imho)
there is no off-hand ready infrastructure to get updates from
replicating node on transaction-by-transaction basis (it currently only
notifies when whole batch is done). For consistency verification (main
reason for this patch) we also don't need invalidations to work, as in
that task we always connect afresh to backup. So I simply only put
relevant TODOs about invalidations for now.

The patch is not very polished but should work.

/reviewed-on nexedi/neoppod!4

Currently, the command "neoctl [arguments] print ids" has the following output:

    last_oid = 0x...
    last_tid = 0x...
    last_ptid = ...

or

    backup_tid = 0x...
    last_tid = 0x...
    last_ptid = ...

depending on whether the cluster is in normal or backup mode.

This is extremely unreadable since the admin is often interested in the time that corresponds to each tid. Now the output is:

    last_oid = 0x...
    last_tid = 0x... (yyyy-mm-dd hh:mm:ss.ssssss)
    last_ptid = ...

or

    backup_tid = 0x... (yyyy-mm-dd hh:mm:ss.ssssss)
    last_tid = 0x... (yyyy-mm-dd hh:mm:ss.ssssss)
    last_ptid = ...

/reviewed-on !2

Similarly to 13911ca3 on the same instance after MariaDB was upgraded to
10.1.17 the following query, even after `OPTIMIZE TABLE obj`, started to execute
very slowly:

    MariaDB [(none)]> SELECT tid FROM neo1.obj WHERE `partition`=5 AND oid=79613 AND tid>268707071353462798 ORDER BY tid LIMIT 1;
    +--------------------+
    | tid                |
    +--------------------+
    | 268707072758797063 |
    +--------------------+
    1 row in set (4.82 sec)

Both explain and analyze says the query will/is using `partition` key but only partially (note key_len is only 10, not 18):

    MariaDB [(none)]> SHOW INDEX FROM neo1.obj;
    +-------+------------+-----------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+---------------+
    | Table | Non_unique | Key_name  | Seq_in_index | Column_name | Collation | Cardinality | Sub_part | Packed | Null | Index_type | Comment | Index_comment |
    +-------+------------+-----------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+---------------+
    | obj   |          0 | PRIMARY   |            1 | partition   | A         |    28755928 |     NULL | NULL   |      | BTREE      |         |               |
    | obj   |          0 | PRIMARY   |            2 | tid         | A         |    28755928 |     NULL | NULL   |      | BTREE      |         |               |
    | obj   |          0 | PRIMARY   |            3 | oid         | A         |    28755928 |     NULL | NULL   |      | BTREE      |         |               |
    | obj   |          0 | partition |            1 | partition   | A         |    28755928 |     NULL | NULL   |      | BTREE      |         |               |
    | obj   |          0 | partition |            2 | oid         | A         |    28755928 |     NULL | NULL   |      | BTREE      |         |               |
    | obj   |          0 | partition |            3 | tid         | A         |    28755928 |     NULL | NULL   |      | BTREE      |         |               |
    | obj   |          1 | data_id   |            1 | data_id     | A         |    28755928 |     NULL | NULL   | YES  | BTREE      |         |               |
    +-------+------------+-----------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+---------------+
    7 rows in set (0.00 sec)

    MariaDB [(none)]> explain SELECT tid FROM neo1.obj WHERE `partition`=5 AND oid=79613 AND tid>268707071353462798 ORDER BY tid LIMIT 1;
    +------+-------------+-------+------+-------------------+-----------+---------+-------------+------+--------------------------+
    | id   | select_type | table | type | possible_keys     | key       | key_len | ref         | rows | Extra                    |
    +------+-------------+-------+------+-------------------+-----------+---------+-------------+------+--------------------------+
    |    1 | SIMPLE      | obj   | ref  | PRIMARY,partition | partition | 10      | const,const |    2 | Using where; Using index |
    +------+-------------+-------+------+-------------------+-----------+---------+-------------+------+--------------------------+
    1 row in set (0.00 sec)

    MariaDB [(none)]> analyze SELECT tid FROM neo1.obj WHERE `partition`=5 AND oid=79613 AND tid>268707071353462798 ORDER BY tid LIMIT 1;
    +------+-------------+-------+------+-------------------+-----------+---------+-------------+------+------------+----------+------------+--------------------------+
    | id   | select_type | table | type | possible_keys     | key       | key_len | ref         | rows | r_rows     | filtered | r_filtered | Extra                    |
    +------+-------------+-------+------+-------------------+-----------+---------+-------------+------+------------+----------+------------+--------------------------+
    |    1 | SIMPLE      | obj   | ref  | PRIMARY,partition | partition | 10      | const,const |    2 | 9741121.00 |   100.00 |       0.00 | Using where; Using index |
    +------+-------------+-------+------+-------------------+-----------+---------+-------------+------+------------+----------+------------+--------------------------+
    1 row in set (4.93 sec)

By explicitly forcing (partition, oid, tid) index usage which is precisely designed to serve this and similar queries can avoid the query from being slow:

    MariaDB [(none)]> analyze SELECT tid FROM neo1.obj FORCE INDEX(`partition`) WHERE `partition`=5 AND oid=79613 AND tid>268707071353462798 ORDER BY tid LIMIT 1;
    +------+-------------+-------+-------+---------------+-----------+---------+------+------+--------+----------+------------+--------------------------+
    | id   | select_type | table | type  | possible_keys | key       | key_len | ref  | rows | r_rows | filtered | r_filtered | Extra                    |
    +------+-------------+-------+-------+---------------+-----------+---------+------+------+--------+----------+------------+--------------------------+
    |    1 | SIMPLE      | obj   | range | partition     | partition | 18      | NULL |    2 |   1.00 |   100.00 |     100.00 | Using where; Using index |
    +------+-------------+-------+-------+---------------+-----------+---------+------+------+--------+----------+------------+--------------------------+
    1 row in set (0.00 sec)

/cc @jm, @vpelltier, @Tyagov

/reviewed-on !1

Many patches have been merged upstream :)

A notable change is that lastTransaction() does not ping the master anymore
(but it still causes a connection to the master if the client is disconnected).

After partitions were dropped with TokuDB, we had a case where MariaDB 10.1.14
stopped using the most appropriate index.

MariaDB [neo0]> explain SELECT DISTINCT data_id FROM obj WHERE `partition`=5;
+------+-------------+-------+-------+-------------------+---------+---------+------+------+---------------------------------------+
| id   | select_type | table | type  | possible_keys     | key     | key_len | ref  | rows | Extra                                 |
+------+-------------+-------+-------+-------------------+---------+---------+------+------+---------------------------------------+
|    1 | SIMPLE      | obj   | range | PRIMARY,partition | data_id | 11      | NULL |   10 | Using where; Using index for group-by |
+------+-------------+-------+-------+-------------------+---------+---------+------+------+---------------------------------------+
MariaDB [neo0]> SELECT SQL_NO_CACHE DISTINCT data_id FROM obj WHERE `partition`=5;
Empty set (1 min 51.47 sec)

Expected:

MariaDB [neo1]> explain SELECT DISTINCT data_id FROM obj WHERE `partition`=4;
+------+-------------+-------+------+-------------------+---------+---------+-------+------+------------------------------+
| id   | select_type | table | type | possible_keys     | key     | key_len | ref   | rows | Extra                        |
+------+-------------+-------+------+-------------------+---------+---------+-------+------+------------------------------+
|    1 | SIMPLE      | obj   | ref  | PRIMARY,partition | PRIMARY | 2       | const |    1 | Using where; Using temporary |
+------+-------------+-------+------+-------------------+---------+---------+-------+------+------------------------------+
1 row in set (0.00 sec)
MariaDB [neo1]> SELECT SQL_NO_CACHE DISTINCT data_id FROM obj WHERE `partition`=4;
Empty set (0.00 sec)

Restarting the server or 'OPTIMIZE TABLE obj; ' does not help.

Such issue could prevent the cluster to start due to timeouts, by always going
back to RECOVERING state.

Freeing disk space when a cell is dropped will have to be implemented with care,
not only for performance reasons.

TRUNCATE was chosen for performance reasons, but it's usually done on small
tables, and not for performance-critical operations. TRUNCATE commits
implicitely, so for pt/ttrans in particular, it's certainly slower due to extra
fsyncs to disk.

On the other side, committing too early can corrupt the database if the storage
node is stopped just after. For example, a failure in changePartitionTable()
can cause 'pt' to remain empty.

It slowed down everything but abortFor(), which is not performance critical.

This reverts commit 7aecdada partially.
There seems to be no bug here, because:
- abortFor() is only called upon a notification from the master that a client
  is disconnected,
- and from the same TCP connection, we only receive a LockInformation packet
  if there's still such a transaction on the master side.

The code removed in abortFor() was redundant with abort().

There was a bug that connections were not maintained during a TPC,
which caused transactions to be aborted when the limit was reached.

Given that oids are spreaded evenly over all partitions, and that clients always
write to all cells of each involved partitions, clients would spend their time
reconnecting to storage nodes as soon as the limit is reached. So such feature
really looks counter-productive.

Client nodes ignored the state of the connection to the master node when reading
data from storage, as long as their partition tables were recent enough. This
way, they were able to finish read-only transactions even if they could't reach
the master, which could be useful for high availability. The downside is that
the master node ignored that their node ids were still used, which causes "uuid"
conflicts when reallocating them.

Rejected solutions:
- An unused NEO Storage should not insist in staying connected to master node.
- Reverting to big random node identifiers is a lot of work and it would make
  debugging annoying (see commit 23fad3af).
- Always increasing node ids could have been a simple solution if we accepted
  that the cluster dies after that all 2^24 possible ids were allocated.

Given that reading from storage without being connected to the master can only
be useful to finish the current transaction (because we always ping the master
at the beginning of every transaction), keeping such feature is not worth the
effort.

This commit fixes id conflicts in a very simple way, by clearing the partition
table upon primary node failure, which forces reconnection to the master before
querying any storage node. In such case, we raise a special exception that will
cause the transaction to be restarted, so that the user does not get errors for
temporary connection failures.

Currently, another argument not to lock is that we would not be able to test
incremental resolution anymore. We can think about this again when deadlock
resolution is implemented.

The following error was reported on a client node:

    #0x0000 Error                   < None (2001:...:2051)
    1 (Retry Later)
    connection closed for <MTClientConnection(uuid=None, address=2001:...:2051, handler=PrimaryNotificationsHandler, closed, client) at 7f1ea7c42f90>
    Event Manager:
    connection started for <MTClientConnection(uuid=None, address=2001:...:2051, handler=PrimaryNotificationsHandler, fd=13, on_close=onConnectionClosed, connecting, client) at 7f1ea7c25c10>
    #0x0000 RequestIdentification          > None (2001:...:2051)
      Readers: []
      Writers: []
      Connections:
        13: <MTClientConnection(uuid=None, address=2001:...:2051, handler=PrimaryNotificationsHandler, fd=13, on_close=onConnectionClosed, connecting, client) at 7f1ea7c25c10> (pending=False)
    Node manager : 1 nodes
    * None |   MASTER | 2001:...:2051 | UNKNOWN
    <ClientCache history_size=0 oid_count=0 size=0 time=0 queue_length=[0] (life_time=10000 max_history_size=100000 max_size=20971520)>
    poll raised, retrying
    Traceback (most recent call last):
      File "neo/lib/threaded_app.py", line 93, in _run
        poll(1)
      File "neo/lib/event.py", line 134, in poll
        self._poll(0)
      File "neo/lib/event.py", line 164, in _poll
        conn = self.connection_dict[fd]
    KeyError: 13

which means that:
- while the poll thread is getting a (13, EPOLLIN) event because it is
  closed (aborted by the master)
- another thread processes the error packet, by closing it in
  PrimaryBootstrapHandler.notReady
- next, the poll thread resumes the execution of EpollEventManager._poll
  and fails to find fd=13 in self.connection_dict

So here, we have a race condition between epoll_wait and any further use
of connection_dict to map returned fds.

However, what commit a4731a0c does to handle
the case of fd reallocation only works for mono-threaded applications.
In EPOLLIN, wrapping 'self.connection_dict[fd]' the same way as for other
events is not enough. For example:
- case 1:
  - thread 1: epoll returns fd=13
  - thread 2: close(13)
  - thread 2: open(13)
  - thread 1: self.connection_dict[13] does not match
              but this would be handled by the 'unregistered' list
- case 2:
  - thread 1: reset 'unregistered'
  - thread 2: close(13)
  - thread 2: open(13)
  - thread 1: epoll returns fd=13
  - thread 1: self.connection_dict[13] matches
              but it would be wrongly ignored by 'unregistered'
- case 3:
  - thread 1: about to call readable/writable/onTimeout on a connection
  - thread 2: this connection is closed
  - thread 1: readable/writable/onTimeout wrongly called on a closed connection

We could protect _poll() with a lock, and make unregister() use wakeup() so
that it gets a chance to acquire it, but that causes threaded tests to deadlock
(continuing in this direction seems too complicated).

So we have to deal with the fact that there can be race conditions at any time
and there's no way to make 'connection_dict' match exactly what epoll returns.
We solve this by preventing fd reallocation inside _poll(), which is fortunately
possible with sockets, using 'shutdown': the closing of fds is delayed.

For above case 3, readable/writable/onTimeout for MTClientConnection are also
changed to test whether the connection is still open while it has the lock.
Just for safety, we do the same for 'process'.

At last, another kind of race condition that this commit also fixes concerns
the use of itervalues() on EventManager.connection_dict.

This is a preliminary commit, without any functional change,
just to make the next one easier to review.

We already have logs when a connection fails,
and ask() raises ConnectionClosed if the connection is closed.