Distributed, redundant and transactional storage for ZODB
- How to use
- Commercial Support
- Change History
- 1.10 (2018-07-16)
- 1.9 (2018-03-13)
- 1.8.1 (2017-11-07)
- 1.8 (2017-07-04)
- 1.7.1 (2017-01-18)
- 1.7.0 (2016-12-19)
- 1.6.3 (2016-06-15)
- 1.6.2 (2016-03-09)
- 1.6.1 (2016-01-25)
- 1.6 (2015-12-02)
- 1.5.1 (2015-10-26)
- 1.5 (2015-10-05)
- 1.4 (2015-07-13)
- 1.3 (2015-01-13)
- 1.2 (2014-07-30)
- 1.1 (2014-01-07)
- 1.0 (2012-08-28)
- 0.10.1 (2012-03-13)
- 0.10 (2011-10-17)
- 0.9.2 (2011-10-17)
- 0.9.1 (2011-09-24)
- 0.9 (2011-09-12)
NEO is a distributed, redundant and scalable implementation of ZODB API. NEO stands for Nexedi Enterprise Object.
A NEO cluster is composed of the following types of nodes:
“master” nodes (mandatory, 1 or more)
Takes care of transactionality. Only one master node is really active (the active master node is called “primary master”) at any given time, extra masters are spares (they are called “secondary masters”).
“storage” nodes (mandatory, 1 or more)
Stores data, preserving history. All available storage nodes are in use simultaneously. This offers redundancy and data distribution. Available backends: MySQL (InnoDB, RocksDB or TokuDB), SQLite
“admin” nodes (mandatory for startup, optional after)
Accepts commands from neoctl tool and transmits them to the primary master, and monitors cluster state.
Well… Something needing to store/load data in a NEO cluster.
ZODB API is fully implemented except:
- pack: only old revisions of objects are removed (it should be possible to use zc.zodbdgc for garbage collection)
- blobs: not implemented (not considered yet)
Any ZODB like FileStorage can be converted to NEO instantaneously, which means the database is operational before all data are imported. There’s also a tool to convert back to FileStorage.
For more detailed information about features related to scalability, see the Architecture and Characteristics section of https://neo.nexedi.com/.
- Linux 2.6 or later
- Python 2.7.x (2.7.9 or later for SSL support)
- For storage nodes using MySQL backend:
- For client nodes: ZODB 3.10.x or later
NEO can be installed like any other egg (see setup.py). Or you can simply make neo directory available for Python to import (for example, by adding its container directory to the PYTHONPATH environment variable).
Write a neo.conf file like the example provided. If you use MySQL, you’ll also need create 1 database per storage node.
Start all required nodes:
$ neomaster -f neo.conf $ neostorage -f neo.conf -s storage1 $ neostorage -f neo.conf -s storage2 $ neoadmin -f neo.conf
Tell the cluster to initialize storage nodes:
$ neoctl -a <admin> start
Clients can connect when the cluster is in RUNNING state:
$ neoctl -a <admin> print cluster RUNNING
See importer.conf file to import an existing database, or neoctl command for more administrative tasks.
Alternatively, you can use neosimple command to quickly setup a cluster for testing.
First make sure Python can import ‘neo.client’ package.
Edit your zope.conf, add a neo import and edit the zodb_db section by replacing its filestorage subsection by a NEOStorage one. It should look like:
%import neo.client <zodb_db main> <NEOStorage> master_nodes 127.0.0.1:10000 name <cluster name> </NEOStorage> mount-point / </zodb_db>
Just create the storage object and play with it:
from neo.client.Storage import Storage s = Storage(master_nodes="127.0.0.1:10010", name="main") ...
“name” and “master_nodes” parameters have the same meaning as in configuration file.
Before shutting down NEO, all clients like Zope instances should be stopped, so that cluster become idle. This is required for multi-DB setups, to prevent critical failures in second phase of TPC.
A cluster (i.e. masters+storages+admin) can be stopped gracefully by putting it in STOPPING state using neoctl:
neoctl -a <admin> set cluster STOPPING
This can also be done manually, which helps if your cluster is in bad state:
- Stop all master nodes first with a SIGINT or SIGTERM, so that storage nodes don’t become in OUT_OF_DATE state.
- Next stop remaining nodes with a SIGINT or SIGTERM.
This is the recommanded way to backup a NEO cluster. Once a cluster with appropriate upstream_cluster & upstream_masters configuration is started, you can switch it into backup mode using:
neoctl -a <admin> set cluster STARTING_BACKUP
It remembers it is in such mode when it is stopped, and it can be put back into normal mode (RUNNING) by setting it into STOPPING_BACKUP state.
Packs are currently not replicated, which means packing should always be done up to a TID that is already fully replicated, so that the backup cluster has a full history (and not random holes).
In addition to any external solution like OpenVPN, NEO has builtin SSL support to authenticate and encrypt communications between nodes.
All commands and configuration files have options to specify a CA certificate, the node certificate and the node private key. A certificate can be shared by several nodes.
NEO always uses the latest SSL protocol supported by the Python interpreter, without fallback to older versions. A “SSL: WRONG_VERSION_NUMBER” error means that a node runs in an older environment (Python + OpenSSL) than others.
Note also that you can’t mix non-SSL nodes and SSL nodes, even between a upstream cluster and a backup one. In doing so, connections can get stuck, or fail with malformed packets or SSL handshake errors.
NEO has no built-in deployment features such as process daemonization. We use supervisor with configuration like below:
[group:neo] programs=master_01,storage_01,admin [program:storage_01] priority=10 command=neostorage -s storage_01 -f /neo/neo.conf [program:master_01] priority=20 command=neomaster -s master_01 -f /neo/neo.conf [program:admin] priority=20 command=neoadmin -s admin -f /neo/neo.conf
Developers interested in NEO may refer to NEO Web site and subscribe to following mailing lists:
Automated test results are published at https://www.erp5.com/quality/integration/P-ERP5.Com.Unit%20Tests/Base_viewListMode?proxy_form_id=WebSection_viewERP5UnitTestForm&proxy_field_id=listbox&proxy_field_selection_name=WebSection_viewERP5UnitTestForm_listbox_selection&reset=1&listbox_title=NEO-%25
A important performance improvement is that the replication now remembers where it was interrupted: a storage node that gets disconnected for a short time now gets fully operational quite instantaneously because it only has to replicate the new data. Before, the time to recover depended on the size of the DB, just to verify that most of the data are already transferred.
As a small optimization, an empty transaction extension is now serialized with an empty string.
The above 2 changes required a bump of the protocol version, as well as an upgrade of the storage format. Once upgraded (this is done automatically as usual), databases can’t be opened anymore by older versions of NEO.
Other general changes:
- Add support for custom compression levels.
- Maximize resiliency by taking into account the topology of storage nodes.
- Fix a few issues with ZODB5. Note however that merging several DB with the Importer backend only works if they were only used with ZODB < 5.
- Automatically discard feeding cells that get out-of-date.
- Fix partial import from a source storage.
- Store uncompressed if compressed size is equal.
- Fixed v1.9 code that sped up the replication by sending bigger network packets.
- Fix replication of creation undone.
- Stop logging ‘Abort TXN’ for txn that have been locked.
- Clarify log about data deletion of discarded cells.
- Fix replication of big OIDs (> 16M).
- Do not full-scan for duplicates of big OIDs if deduplication is disabled.
- Fix remaining places where a server disconnection was not catched.
- Fix indexes of upgraded databases.
- Fetch and process the data to import in a separate process. It is even usually free to use the best compression level.
- New option to write back new transactions to the source database. See ‘importer.conf’ for more information.
- Give a title to the ‘import’ and ‘writeback’ subprocesses, if the ‘setproctitle’ egg is installed.
- Log when the transaction index for FileStorage DB is built.
- Open imported database in read-only whenever possible.
- Do not trigger speedupFileStorageTxnLookup uselessly.
- Do not checksum data twice.
- Fix NameError when recovering during tpc_finish.
A lot of performance improvements have been done on storage nodes for this release, and some of them required changes in the storage format. In short, the migration is done automatically, but you may want to read UPGRADE notes for more details.
- Speed up replication by sending bigger network packets, and by not getting object next_serial for nothing.
- Speed up reads by indexing ‘obj’ primarily by ‘oid’ (instead of ‘tid’).
- Optimize storage layout of raw data for replication.
Other storage changes:
- Disable data deduplication by default. –dedup option added.
- importer: do not crash if a backup cluster tries to replicate.
- importer: preserve ‘packed’ flag.
- Fix possible failure when reading data in a backup cluster with replicas.
- Fix generation of TID.
- Fix resumption of backup replication (internal or not).
- New ‘cache-size’ Storage option.
- Cache hit/miss statistics.
- Fix accounting of cache size.
- Preserve ‘packed’ flag on iteration.
- At startup, or after nodes are back, full load balancing could be prevented until some data are written.
- neolog: –from option now also tries to parse with dateutil.
- neolog: add support for xz-compressed logs, using external xzcat commands.
- Add support for OpenSSL >= 1.1.
- storage: fix possible crash when delaying replication requests.
- mysql: fix performance issues on read operations by using more index hints.
This release mainly stabilizes NEO when it is used with several storage nodes, fixing many race conditions involving events like transactional operations (read/write, conflict resolution…), replication, partition table tweaking, and all kinds of failures (node crashes, network cuts…). This includes a rework of conflict resolution, to implement the long-awaited deadlock avoidance (it was a limitation caused by object-level locking).
Similarly, having spare master nodes is not an experimental feature anymore: the election (of the primary master) has been reimplemented, and it now happens during the RECOVERING phase. This comes with a change about node states: BROKEN/HIDDEN/UNKNOWN are removed, DOWN is renamed into UNKNOWN, and TEMPORARILY_DOWN into DOWN.
And still for more resiliency, the new algorithm to tweak the partition table is better at minimizing the amount of replication, and it does not discard readable cells too quickly anymore: a partition can now have multiple FEEDING cells, to avoid going below the wanted level of replication.
- Packet timeouts have been removed. TCP keepalives are used instead of applicative pings.
- Connection handshake between nodes is reviewed to make sure that they speak the same protocol before doing anything else, and report clearer error messages otherwise. A dangerous bug was that there was no protocol version check between neoctl and the admin node.
- Proper handling of incoming packets for closed/aborted connections.
- An exception while processing an answer could leave the handler switcher in the bad state.
- In STOPPING cluster state, really wait for all transactions to be finished.
- Several issues when undoing transactions with conflict resolutions have been fixed.
- Delayed connection acceptation when the storage node is ready.
- Added support for zodburi.
- Fix load error during conflict resolution in case of late invalidation.
- Do not wait tpc_vote to start resolving conflicts.
- Fix harmless ‘unexpected … AnswerRequestIdentification’ exceptions.
- New –disable-drop-partitions option, which is useful for big databases because the current code to delete data of discarded cells is inefficient (this option should disappear in the future).
- Prevent 2 nodes from working with the same database.
- Discard answers from aborted replications. In some cases, this led to data corruption or crashes.
- MySQL backend:
- Added support for RocksDB.
- Do not flood logs when retrying to connect non-stop.
- Do not retry a failing query forever.
- By default, do not retry to connect to the server automatically.
- neolog: new –decompress option.
- neolog: new option to hide the node column.
- neoctl: make the identification of the primary master easier with ‘print node’.
- A lot of improvements for developers and debugging.
- Fixed possibly wrong knowledge of cells’ backup_tid when resuming backup. In such case, ‘neoctl print ids’ gave false impression that the backup cluster was up-to-date. This also resulted in an inconsistent database when leaving backup mode before that the issue resolved by itself.
- Storage nodes now select the partition which is furthest behind. Previous criterion was such that in case of high upstream activity, the backup could even be stuck looping on a subset of partitions.
- Fixed replication of unfinished imported transactions.
- Fixed abort before vote, to free the storage space used by the transaction. A new ‘prune_orphan’ neoctl command was added to delete unreferenced raw data in the database.
- Removed short storage option -R to reset the db. Help is reworded to clarify that –reset exits once done.
- The application receiving buffer size has been increased. This speeds up transfer of big packets.
- The master raised AttributeError at exit during recovery.
- At startup, the importer storage backend connected twice to the destination database.
- Identification issues, mainly caused by id conflicts, are fixed:
- Storage nodes now only accept clients that are known by the master.
- When reconnecting to a master, a client get a new id if the previous id is already reallocated to another client.
- The consequences were either crashes or clients being unable to connect.
- Added support for the latest versions of ZODB (4.4.4 & 5.0.1). 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).
- A cluster in BACKUPING state can now serve regular clients in read-only mode. But without invalidation yet, so clients must reconnect whenever they want to see newer data.
- Fixed crash of client nodes (including backup master) while trying to process notifications before complete initialization, instead of ignoring them.
- Fix race condition leading to invalid mapping between internal connection objects and their file descriptors. This resulted in KeyError exceptions.
- Fix item eviction from cache, which could break loading from storage.
- Better exception handling in tpc_abort.
- Do not limit the number of open connections to storage nodes.
- Fix crash when a client loses connection to the master just before voting.
- MySQL: Force index for a few queries. Unfortunately, this is not perfect because sometimes MySQL still ignores our hints.
- MySQL: Do not use unsafe TRUNCATE statement.
- Make ‘neoctl print ids’ display time of TIDs.
- Various neoctl/neolog formatting improvements/fixes.
- Plus a few other changes for debugging and developers, as well as small optimizations.
- Added support for ZODB 4.x
- Clients are now able to recover from failures during tpc_finish when the transaction got successfully committed.
- Other fixes related to node disconnection:
- storage: fix crash when a client disconnects just after it requested to finish a transaction
- storage: fix crash when trying to replicate from an unreachable node
- master: do never abort a prepared transaction (for example, a client disconnecting during tpc_finish could cause a crash)
- client: fix invalidation issues when reconnecting to the master
- fix abort for storages where only current serials were checked
- fix the count of history items in the cache
- neoctl: better error message when connection to admin fails
- storage: switch to a maintained fork of MySQL-python (mysqlclient)
- storage: for better performance, the backend commit after an unlocked transaction is deferred by 1 second, with the hope it’s merged by a subsequent commit (in case of a crash, the transaction is unlocked again), so there are only 2 commits per transaction during high activity
- client: optimize cache by not keeping items with counter=0 in history queue
- client: fix possible assertion failure on load in case of a late invalidation
NEO repository has moved to https://lab.nexedi.com/nexedi/neoppod.git
- client: fix spurious connection timeouts
- client: add cache stats to information dumped on SIGRTMIN+2
- storage: when using the Importer backend, allow truncation after the last tid to import, during or after the import
- neoctl: don’t print ‘None’ on successful check/truncate commands
- neolog: fix crash on unknown packets
- plus a few other changes for debugging and developers
This release has changes in storage format. The upgrade is done automatically, but only if the cluster was stopped cleanly: see UPGRADE notes for more information.
NEO did not ensure that all data and metadata were written on disk before tpc_finish, and it was for example vulnerable to ENOSPC errors. In order to minimize the risk of failures during tpc_finish, the writing of metadata to temporary tables is now done in tpc_vote. See commit 7eb7cf1 for more information about possible changes on performance side.
This change comes with a new algorithm to verify unfinished data, which also fixes a bug discarding transactions with objects for which readCurrent was called.
The RECOVERING/VERIFYING phases, as well as transitions from/to other states, have been completely reviewed, to fix many bugs:
- Possible corruption of partition table.
- The cluster could be stuck in RECOVERING or VERIFYING state.
- The probability to have cells out-of-date when restarting several storage nodes simultaneously has been reduced.
- During recovery, a newly elected master now always waits all the storage nodes with readable cells to be pending, in order to avoid a split of the database.
- The last tid/oid could be wrong in several cases, for example after transactions are recovered during VERIFYING phase.
neoctl gets a new command to truncate the database at an arbitrary TID. Internally, NEO was already able to truncate the database, because this was necessary to make the database consistent when leaving the backup mode. However, there were several bugs that caused the database to be partially truncated:
- The master now first stores persistently the decision to truncate, so that it can recover from any kind of connection failure.
- The cluster goes back to RUNNING state only after an acknowledgment from all storage nodes (including those without any readable cell) that they truncated.
- As a workaround to fix holes if replication is interrupted after new data is committed, outdated cells always restart to replicate from the beginning.
- The deletion of partial transactions during verification didn’t try to free the associated raw data.
- The MySQL backend didn’t drop the ‘bigdata’ table when erasing the database.
Handshaking SSL connections could be stuck when they’re aborted.
‘neoctl print ids’ displays a new value in backup mode: the highest common TID up to which all readable cells have replicated, i.e. the TID at which the database would be truncated when leaving the backup mode.
Several bugs and performance issues have been fixed in this release, mainly in the storage node.
- Importer storage backend:
- Fix retrieval of an object from ZODB when next serial in NEO.
- Fix crash of storage nodes when a transaction is aborted.
- Faster resumption when many transactions have already been imported to MySQL.
- MySQL storage backend:
- Refuse to start if max_allowed_packet is too small.
- Faster commit of transaction metadata.
- Replication & checking of replicas:
- Fix crash when a corruption is found while checking TIDs. 2 other issues remain unfixed: see BUGS.rst file.
- Speed up checking of replicas, at the cost of storage nodes being less responsive to other events.
- The master wrongly sent invalidations for objects on which only readCurrent was called, which caused invalid entries in client caches, or assertion failures in Connection._setstate_noncurrent.
In this version, the connectivity between nodes has been greatly improved:
- Added SSL support.
- IPv4 & IPv6 can be mixed: some nodes can have an IPv4 binding address, whereas other listen on IPv6.
- Version 1.4 fixed several cases where nodes could reconnect too quickly, using 100% CPU and flooding logs. This is now fixed completely, for example when a backup storage node was rejected because the upstream cluster was not ready.
- Tickless poll loop, for lower latency and CPU usage: nodes don’t wake up every second anymore to check if a timeout has expired.
- Connections could be wrongly processed before being polled (for reading or writing). This happened if a file descriptor number was reallocated by the kernel for a connection, just after a connection was closed.
Other changes are:
- IStorage: history() did not wait the oid to be unlocked. This means that the latest version of an object could be missing from the result.
- Log files can now be specified in configuration files.
- ~(user) construction are expanded for all paths in configuration (file or command line). This does not concern non-daemon executables like neoctl.
- For neoctl, -l option now logs everything on disk automatically.
- The admin node do not reset anymore the list of known masters from configuration when reconnecting, for consistency with client nodes.
- Code refactoring and improvements to logging and debugging.
- An notable change in the test suite is that the occasional deadlocks that affected threaded tests have been fixed.
This version comes with a change in the SQL tables format, to fix a potential crash of storage nodes when storing values that only differ by the compression flag. See UPGRADE notes if you think your application may be affected by this bug.
- Performance and features:
- ‘Importer’ storage backend has been significantly sped up.
- Support for TokuDB has been added to MySQL storage backend. The engine is still InnoDB by default, and it can be selected via a new ‘neostorage’ option.
- A ‘neomaster’ option has been added to automatically start a new cluster if the number of pending storage nodes is greater than or equal to the specified value.
- Storage crashed when reading empty transactions. We still need to decide
whether NEO should:
- continue to store such transactions;
- ignore them on commit, like other ZODB implementation;
- or fail on commit.
- Storage crashed when a client tries to “steal” the UUID of another client.
- Client could get stuck forever on unreadable cells when not connected to the master.
- Client could only instantiate NEOStorage from the main thread, and the RTMIN+2 signal displayed logs for only 1 NEOStorage. Now, RTMIN+2 & RTMIN+3 are setup when neo.client module is imported.
- Storage crashed when reading empty transactions. We still need to decide whether NEO should:
- Plus fixes and improvements to logging and debugging.
- Version 1.2 added a new ‘Importer’ storage backend but it had 2 bugs.
- An interrupted migration could not be resumed.
- Merging several ZODB only worked if NEO could import all classes used by the application. This has been fixed by repickling without loading any object.
- Logging has been improved for a better integration with the environment:
- RTMIN+1 signal was changed to reopen logs. RTMIN+1 & RTMIN+2 signals, which were previously used for debugging, have been remapped to RTMIN+2 & RTMIN+3
- In Zope, client registers automatically for log rotation (USR2).
- NEO logs are SQLite DB that are not open anymore with a persistent journal, because this is incompatible with the rename+reopen way to rotate logs, and we want to support logrotate.
- ‘neolog’ can now open gzip/bz2 compressed logs transparently.
- ‘neolog’ does not spam the console anymore when piped to a process that exits prematurely.
- MySQL backend has been updated to work with recent MariaDB (>=10).
- 2 ‘neomaster’ command-line options were added to set upstream cluster/masters.
The most important changes in this version are the work about conversion of databases from/to NEO:
A new ‘Importer’ storage backend has been implemented and this is now the recommended way to migrate existing Zope databases. See ‘importer.conf’ example file for more information.
‘neomigrate’ command refused to run since version 1.0
Exported data serials by NEO iterator were wrong. There are still differences with FileStorage:
- NEO always resolves to original serial, to avoid any indirection (which slightly speeds up undo at the expense of a more complex pack code)
- NEO does not make any difference between object deletion and creation undone (data serial always null in storage)
Apart from that, conversion of database back from NEO should be fixed.
Other changes are:
- A warning was added in ‘neo.conf’ about a possible misuse of replicas.
- Compatibility with Python 2.6 has been dropped.
- Support for recent version of SQlite has been added.
- A memory leak has been fixed in replication.
- MySQL backend now fails instead of silently reconnecting if there is any pending change, which could cause data loss.
- Optimization and minor bugfixes.
- Client failed at reconnecting properly to master. It could kill the master (during tpc_finish!) or end up with invalid caches (i.e. possible data corruption). Now, connection to master is even optional between transaction.begin() and tpc_begin, as long as partition table contains up-to-date data.
- Compatibility with ZODB 3.9 has been dropped. Only 3.10.x branch is supported.
- checkCurrentSerialInTransaction was not working.
- Optimization and minor bugfixes.
This version mainly comes with stabilized SQL tables format and efficient backup feature, relying on replication, which has been fully reimplemented:
- It is now incremental, instead of being done on whole partitions. Schema of MySQL tables have been changed in order to optimize storage layout, for good partial replication performance.
- It runs at lowest priority not to degrade performance for client nodes.
- A cluster in the new BACKINGUP state is a client to a normal cluster and all its storage nodes are notified of invalidations and replicate from upstream nodes.
Other changes are:
- Compatibility with Python < 2.6 and ZODB < 3.9 has been dropped.
- Cluster is now automatically started when all storage nodes of UP_TO_DATE cells are available, similarly to mdadm assemble --no-degraded behaviour.
- NEO learned to check replicas, to detect data corruption or bugs during replication. When done on a backup cluster, upstream data is used as reference. This is still limited to data indexes (tid & oid/serial).
- NEO logs now are SQLite DB that always contain all debugging information including exchanged packets. Records are first kept in RAM, at most 16 MB by default, and there are flushed to disk only upon RTMIN signal or any important record. A ‘neolog’ script has been written to help reading such DB.
- Master addresses must be separated by spaces. ‘/’ can’t be used anymore.
- Adding and removing master nodes is now easier: unknown incoming master nodes are now accepted instead of rejected, and nodes can be given a path to a file that maintains a list of known master nodes.
- Node UUIDs have been shortened from 16 to 4 bytes, for better performance and easier debugging.
Also contains code clean-ups and bugfixes.
- Client didn’t limit its memory usage when committing big transactions.
- Master failed to disconnect clients when cluster leaves RUNNING state.
- Storage was unable or slow to process large-sized transactions. This required to change protocol and MySQL tables format.
- NEO learned to store empty values (although it’s useless when managed by a ZODB Connection).
- storage: a specific socket can be given to MySQL backend
- storage: a ConflictError could happen when client is much faster than master
- ‘verbose’ command line option of ‘neomigrate’ did not work
- client: ZODB monkey-patch randomly raised a NameError
- client: method to retrieve history of persistent objects was incompatible with recent ZODB and needlessly asked all storages systematically.
- neoctl: ‘print node’ command (to get list of all nodes) raised an AssertionError.
- ‘neomigrate’ raised a TypeError when converting NEO DB back to FileStorage.
NEO is considered stable enough to replace existing ZEO setups, except that:
- there’s no backup mechanism (aka efficient snapshoting): there’s only replication and underlying MySQL tools
- MySQL tables format may change in the future