5 This document describes the major changes in Ganeti 2.0 compared to
8 The 2.0 version will constitute a rewrite of the 'core' architecture,
9 paving the way for additional features in future 2.x versions.
11 .. contents:: :depth: 3
16 Ganeti 1.2 has many scalability issues and restrictions due to its
17 roots as software for managing small and 'static' clusters.
19 Version 2.0 will attempt to remedy first the scalability issues and
20 then the restrictions.
25 While Ganeti 1.2 is usable, it severely limits the flexibility of the
26 cluster administration and imposes a very rigid model. It has the
27 following main scalability issues:
29 - only one operation at a time on the cluster [#]_
30 - poor handling of node failures in the cluster
31 - mixing hypervisors in a cluster not allowed
33 It also has a number of artificial restrictions, due to historical
36 - fixed number of disks (two) per instance
37 - fixed number of NICs
39 .. [#] Replace disks will release the lock, but this is an exception
40 and not a recommended way to operate
42 The 2.0 version is intended to address some of these problems, and
43 create a more flexible code base for future developments.
45 Among these problems, the single-operation at a time restriction is
46 biggest issue with the current version of Ganeti. It is such a big
47 impediment in operating bigger clusters that many times one is tempted
48 to remove the lock just to do a simple operation like start instance
49 while an OS installation is running.
54 Ganeti 1.2 has a single global lock, which is used for all cluster
55 operations. This has been painful at various times, for example:
57 - It is impossible for two people to efficiently interact with a cluster
58 (for example for debugging) at the same time.
59 - When batch jobs are running it's impossible to do other work (for
60 example failovers/fixes) on a cluster.
62 This poses scalability problems: as clusters grow in node and instance
63 size it's a lot more likely that operations which one could conceive
64 should run in parallel (for example because they happen on different
65 nodes) are actually stalling each other while waiting for the global
66 lock, without a real reason for that to happen.
68 One of the main causes of this global lock (beside the higher
69 difficulty of ensuring data consistency in a more granular lock model)
70 is the fact that currently there is no long-lived process in Ganeti
71 that can coordinate multiple operations. Each command tries to acquire
72 the so called *cmd* lock and when it succeeds, it takes complete
73 ownership of the cluster configuration and state.
75 Other scalability problems are due the design of the DRBD device
76 model, which assumed at its creation a low (one to four) number of
77 instances per node, which is no longer true with today's hardware.
79 Artificial restrictions
80 -----------------------
82 Ganeti 1.2 (and previous versions) have a fixed two-disks, one-NIC per
83 instance model. This is a purely artificial restrictions, but it
84 touches multiple areas (configuration, import/export, command line)
85 that it's more fitted to a major release than a minor one.
90 The fact that each command is a separate process that reads the
91 cluster state, executes the command, and saves the new state is also
92 an issue on big clusters where the configuration data for the cluster
93 begins to be non-trivial in size.
98 In order to solve the scalability problems, a rewrite of the core
99 design of Ganeti is required. While the cluster operations themselves
100 won't change (e.g. start instance will do the same things, the way
101 these operations are scheduled internally will change radically.
103 The new design will change the cluster architecture to:
105 .. image:: arch-2.0.png
107 This differs from the 1.2 architecture by the addition of the master
108 daemon, which will be the only entity to talk to the node daemons.
114 The changes for 2.0 can be split into roughly three areas:
116 - core changes that affect the design of the software
117 - features (or restriction removals) but which do not have a wide
119 - user-level and API-level changes which translate into differences for
120 the operation of the cluster
125 The main changes will be switching from a per-process model to a
126 daemon based model, where the individual gnt-* commands will be
127 clients that talk to this daemon (see `Master daemon`_). This will
128 allow us to get rid of the global cluster lock for most operations,
129 having instead a per-object lock (see `Granular locking`_). Also, the
130 daemon will be able to queue jobs, and this will allow the individual
131 clients to submit jobs without waiting for them to finish, and also
132 see the result of old requests (see `Job Queue`_).
134 Beside these major changes, another 'core' change but that will not be
135 as visible to the users will be changing the model of object attribute
136 storage, and separate that into name spaces (such that an Xen PVM
137 instance will not have the Xen HVM parameters). This will allow future
138 flexibility in defining additional parameters. For more details see
139 `Object parameters`_.
141 The various changes brought in by the master daemon model and the
142 read-write RAPI will require changes to the cluster security; we move
143 away from Twisted and use HTTP(s) for intra- and extra-cluster
144 communications. For more details, see the security document in the
150 In Ganeti 2.0, we will have the following *entities*:
152 - the master daemon (on the master node)
153 - the node daemon (on all nodes)
154 - the command line tools (on the master node)
155 - the RAPI daemon (on the master node)
157 The master-daemon related interaction paths are:
159 - (CLI tools/RAPI daemon) and the master daemon, via the so called
161 - the master daemon and the node daemons, via the node RPC
163 There are also some additional interaction paths for exceptional cases:
165 - CLI tools might access via SSH the nodes (for ``gnt-cluster copyfile``
166 and ``gnt-cluster command``)
167 - master failover is a special case when a non-master node will SSH
168 and do node-RPC calls to the current master
170 The protocol between the master daemon and the node daemons will be
171 changed from (Ganeti 1.2) Twisted PB (perspective broker) to HTTP(S),
172 using a simple PUT/GET of JSON-encoded messages. This is done due to
173 difficulties in working with the Twisted framework and its protocols
174 in a multithreaded environment, which we can overcome by using a
175 simpler stack (see the caveats section).
177 The protocol between the CLI/RAPI and the master daemon will be a
178 custom one (called *LUXI*): on a UNIX socket on the master node, with
179 rights restricted by filesystem permissions, the CLI/RAPI will talk to
180 the master daemon using JSON-encoded messages.
182 The operations supported over this internal protocol will be encoded
183 via a python library that will expose a simple API for its
184 users. Internally, the protocol will simply encode all objects in JSON
185 format and decode them on the receiver side.
187 For more details about the RAPI daemon see `Remote API changes`_, and
188 for the node daemon see `Node daemon changes`_.
193 As described above, the protocol for making requests or queries to the
194 master daemon will be a UNIX-socket based simple RPC of JSON-encoded
197 The choice of UNIX was in order to get rid of the need of
198 authentication and authorisation inside Ganeti; for 2.0, the
199 permissions on the Unix socket itself will determine the access
202 We will have two main classes of operations over this API:
204 - cluster query functions
205 - job related functions
207 The cluster query functions are usually short-duration, and are the
208 equivalent of the ``OP_QUERY_*`` opcodes in Ganeti 1.2 (and they are
209 internally implemented still with these opcodes). The clients are
210 guaranteed to receive the response in a reasonable time via a timeout.
212 The job-related functions will be:
215 - query job (which could also be categorized in the query-functions)
216 - archive job (see the job queue design doc)
217 - wait for job change, which allows a client to wait without polling
219 For more details of the actual operation list, see the `Job Queue`_.
221 Both requests and responses will consist of a JSON-encoded message
222 followed by the ``ETX`` character (ASCII decimal 3), which is not a
223 valid character in JSON messages and thus can serve as a message
224 delimiter. The contents of the messages will be a dictionary with two
228 the name of the method called
230 the arguments to the method, as a list (no keyword arguments allowed)
232 Responses will follow the same format, with the two fields being:
235 a boolean denoting the success of the operation
237 the actual result, or error message in case of failure
239 There are two special value for the result field:
241 - in the case that the operation failed, and this field is a list of
242 length two, the client library will try to interpret is as an
243 exception, the first element being the exception type and the second
244 one the actual exception arguments; this will allow a simple method of
245 passing Ganeti-related exception across the interface
246 - for the *WaitForChange* call (that waits on the server for a job to
247 change status), if the result is equal to ``nochange`` instead of the
248 usual result for this call (a list of changes), then the library will
249 internally retry the call; this is done in order to differentiate
250 internally between master daemon hung and job simply not changed
252 Users of the API that don't use the provided python library should
253 take care of the above two cases.
256 Master daemon implementation
257 ++++++++++++++++++++++++++++
259 The daemon will be based around a main I/O thread that will wait for
260 new requests from the clients, and that does the setup/shutdown of the
261 other thread (pools).
263 There will two other classes of threads in the daemon:
265 - job processing threads, part of a thread pool, and which are
266 long-lived, started at daemon startup and terminated only at shutdown
268 - client I/O threads, which are the ones that talk the local protocol
269 (LUXI) to the clients, and are short-lived
271 Master startup/failover
272 +++++++++++++++++++++++
274 In Ganeti 1.x there is no protection against failing over the master
275 to a node with stale configuration. In effect, the responsibility of
276 correct failovers falls on the admin. This is true both for the new
277 master and for when an old, offline master startup.
279 Since in 2.x we are extending the cluster state to cover the job queue
280 and have a daemon that will execute by itself the job queue, we want
281 to have more resilience for the master role.
283 The following algorithm will happen whenever a node is ready to
284 transition to the master role, either at startup time or at node
287 #. read the configuration file and parse the node list
290 #. query all the nodes and make sure we obtain an agreement via
291 a quorum of at least half plus one nodes for the following:
293 - we have the latest configuration and job list (as
294 determined by the serial number on the configuration and
295 highest job ID on the job queue)
297 - if we are not failing over (but just starting), the
298 quorum agrees that we are the designated master
300 - if any of the above is false, we prevent the current operation
301 (i.e. we don't become the master)
303 #. at this point, the node transitions to the master role
305 #. for all the in-progress jobs, mark them as failed, with
306 reason unknown or something similar (master failed, etc.)
308 Since due to exceptional conditions we could have a situation in which
309 no node can become the master due to inconsistent data, we will have
310 an override switch for the master daemon startup that will assume the
311 current node has the right data and will replicate all the
312 configuration files to the other nodes.
314 **Note**: the above algorithm is by no means an election algorithm; it
315 is a *confirmation* of the master role currently held by a node.
320 The logging system will be switched completely to the standard python
321 logging module; currently it's logging-based, but exposes a different
322 API, which is just overhead. As such, the code will be switched over
323 to standard logging calls, and only the setup will be custom.
325 With this change, we will remove the separate debug/info/error logs,
326 and instead have always one logfile per daemon model:
328 - master-daemon.log for the master daemon
329 - node-daemon.log for the node daemon (this is the same as in 1.2)
330 - rapi-daemon.log for the RAPI daemon logs
331 - rapi-access.log, an additional log file for the RAPI that will be
332 in the standard HTTP log format for possible parsing by other tools
334 Since the :term:`watcher` will only submit jobs to the master for
335 startup of the instances, its log file will contain less information
336 than before, mainly that it will start the instance, but not the
342 The only change to the node daemon is that, since we need better
343 concurrency, we don't process the inter-node RPC calls in the node
344 daemon itself, but we fork and process each request in a separate
347 Since we don't have many calls, and we only fork (not exec), the
348 overhead should be minimal.
353 A discussed alternative is to keep the current individual processes
354 touching the cluster configuration model. The reasons we have not
355 chosen this approach is:
357 - the speed of reading and unserializing the cluster state
358 today is not small enough that we can ignore it; the addition of
359 the job queue will make the startup cost even higher. While this
360 runtime cost is low, it can be on the order of a few seconds on
361 bigger clusters, which for very quick commands is comparable to
362 the actual duration of the computation itself
364 - individual commands would make it harder to implement a
365 fire-and-forget job request, along the lines "start this
366 instance but do not wait for it to finish"; it would require a
367 model of backgrounding the operation and other things that are
368 much better served by a daemon-based model
370 Another area of discussion is moving away from Twisted in this new
371 implementation. While Twisted has its advantages, there are also many
372 disadvantages to using it:
374 - first and foremost, it's not a library, but a framework; thus, if
375 you use twisted, all the code needs to be 'twiste-ized' and written
376 in an asynchronous manner, using deferreds; while this method works,
377 it's not a common way to code and it requires that the entire process
378 workflow is based around a single *reactor* (Twisted name for a main
380 - the more advanced granular locking that we want to implement would
381 require, if written in the async-manner, deep integration with the
382 Twisted stack, to such an extend that business-logic is inseparable
383 from the protocol coding; we felt that this is an unreasonable
384 request, and that a good protocol library should allow complete
385 separation of low-level protocol calls and business logic; by
386 comparison, the threaded approach combined with HTTPs protocol
387 required (for the first iteration) absolutely no changes from the 1.2
388 code, and later changes for optimizing the inter-node RPC calls
389 required just syntactic changes (e.g. ``rpc.call_...`` to
390 ``self.rpc.call_...``)
392 Another issue is with the Twisted API stability - during the Ganeti
393 1.x lifetime, we had to to implement many times workarounds to changes
394 in the Twisted version, so that for example 1.2 is able to use both
397 In the end, since we already had an HTTP server library for the RAPI,
398 we just reused that for inter-node communication.
404 We want to make sure that multiple operations can run in parallel on a
405 Ganeti Cluster. In order for this to happen we need to make sure
406 concurrently run operations don't step on each other toes and break the
409 This design addresses how we are going to deal with locking so that:
411 - we preserve data coherency
412 - we prevent deadlocks
413 - we prevent job starvation
415 Reaching the maximum possible parallelism is a Non-Goal. We have
416 identified a set of operations that are currently bottlenecks and need
417 to be parallelised and have worked on those. In the future it will be
418 possible to address other needs, thus making the cluster more and more
419 parallel one step at a time.
421 This section only talks about parallelising Ganeti level operations, aka
422 Logical Units, and the locking needed for that. Any other
423 synchronization lock needed internally by the code is outside its scope.
428 The proposed library has these features:
430 - internally managing all the locks, making the implementation
431 transparent from their usage
432 - automatically grabbing multiple locks in the right order (avoid
434 - ability to transparently handle conversion to more granularity
435 - support asynchronous operation (future goal)
437 Locking will be valid only on the master node and will not be a
438 distributed operation. Therefore, in case of master failure, the
439 operations currently running will be aborted and the locks will be
440 lost; it remains to the administrator to cleanup (if needed) the
441 operation result (e.g. make sure an instance is either installed
442 correctly or removed).
444 A corollary of this is that a master-failover operation with both
445 masters alive needs to happen while no operations are running, and
446 therefore no locks are held.
448 All the locks will be represented by objects (like
449 ``lockings.SharedLock``), and the individual locks for each object
450 will be created at initialisation time, from the config file.
452 The API will have a way to grab one or more than one locks at the same
453 time. Any attempt to grab a lock while already holding one in the wrong
454 order will be checked for, and fail.
460 At the first stage we have decided to provide the following locks:
462 - One "config file" lock
463 - One lock per node in the cluster
464 - One lock per instance in the cluster
466 All the instance locks will need to be taken before the node locks, and
467 the node locks before the config lock. Locks will need to be acquired at
468 the same time for multiple instances and nodes, and internal ordering
469 will be dealt within the locking library, which, for simplicity, will
470 just use alphabetical order.
472 Each lock has the following three possible statuses:
474 - unlocked (anyone can grab the lock)
475 - shared (anyone can grab/have the lock but only in shared mode)
476 - exclusive (no one else can grab/have the lock)
478 Handling conversion to more granularity
479 +++++++++++++++++++++++++++++++++++++++
481 In order to convert to a more granular approach transparently each time
482 we split a lock into more we'll create a "metalock", which will depend
483 on those sub-locks and live for the time necessary for all the code to
484 convert (or forever, in some conditions). When a metalock exists all
485 converted code must acquire it in shared mode, so it can run
486 concurrently, but still be exclusive with old code, which acquires it
489 In the beginning the only such lock will be what replaces the current
490 "command" lock, and will acquire all the locks in the system, before
491 proceeding. This lock will be called the "Big Ganeti Lock" because
492 holding that one will avoid any other concurrent Ganeti operations.
494 We might also want to devise more metalocks (eg. all nodes, all
495 nodes+config) in order to make it easier for some parts of the code to
496 acquire what it needs without specifying it explicitly.
498 In the future things like the node locks could become metalocks, should
499 we decide to split them into an even more fine grained approach, but
500 this will probably be only after the first 2.0 version has been
503 Adding/Removing locks
504 +++++++++++++++++++++
506 When a new instance or a new node is created an associated lock must be
507 added to the list. The relevant code will need to inform the locking
508 library of such a change.
510 This needs to be compatible with every other lock in the system,
511 especially metalocks that guarantee to grab sets of resources without
512 specifying them explicitly. The implementation of this will be handled
513 in the locking library itself.
515 When instances or nodes disappear from the cluster the relevant locks
516 must be removed. This is easier than adding new elements, as the code
517 which removes them must own them exclusively already, and thus deals
518 with metalocks exactly as normal code acquiring those locks. Any
519 operation queuing on a removed lock will fail after its removal.
521 Asynchronous operations
522 +++++++++++++++++++++++
524 For the first version the locking library will only export synchronous
525 operations, which will block till the needed lock are held, and only
526 fail if the request is impossible or somehow erroneous.
528 In the future we may want to implement different types of asynchronous
531 - try to acquire this lock set and fail if not possible
532 - try to acquire one of these lock sets and return the first one you
533 were able to get (or after a timeout) (select/poll like)
535 These operations can be used to prioritize operations based on available
536 locks, rather than making them just blindly queue for acquiring them.
537 The inherent risk, though, is that any code using the first operation,
538 or setting a timeout for the second one, is susceptible to starvation
539 and thus may never be able to get the required locks and complete
540 certain tasks. Considering this providing/using these operations should
541 not be among our first priorities.
546 For the first version of this code we'll convert each Logical Unit to
547 acquire/release the locks it needs, so locking will be at the Logical
548 Unit level. In the future we may want to split logical units in
549 independent "tasklets" with their own locking requirements. A different
550 design doc (or mini design doc) will cover the move from Logical Units
556 In general when acquiring locks we should use a code path equivalent
566 This makes sure we release all locks, and avoid possible deadlocks. Of
567 course extra care must be used not to leave, if possible locked
568 structures in an unusable state. Note that with Python 2.5 a simpler
569 syntax will be possible, but we want to keep compatibility with Python
570 2.4 so the new constructs should not be used.
572 In order to avoid this extra indentation and code changes everywhere in
573 the Logical Units code, we decided to allow LUs to declare locks, and
574 then execute their code with their locks acquired. In the new world LUs
575 are called like this::
577 # user passed names are expanded to the internal lock/resource name,
578 # then known needed locks are declared
580 ... some locking/adding of locks may happen ...
581 # late declaration of locks for one level: this is useful because sometimes
582 # we can't know which resource we need before locking the previous level
583 lu.DeclareLocks() # for each level (cluster, instance, node)
584 ... more locking/adding of locks can happen ...
585 # these functions are called with the proper locks held
588 ... locks declared for removal are removed, all acquired locks released ...
590 The Processor and the LogicalUnit class will contain exact documentation
591 on how locks are supposed to be declared.
596 This library will provide an easy upgrade path to bring all the code to
597 granular locking without breaking everything, and it will also guarantee
598 against a lot of common errors. Code switching from the old "lock
599 everything" lock to the new system, though, needs to be carefully
600 scrutinised to be sure it is really acquiring all the necessary locks,
601 and none has been overlooked or forgotten.
603 The code can contain other locks outside of this library, to synchronise
604 other threaded code (eg for the job queue) but in general these should
605 be leaf locks or carefully structured non-leaf ones, to avoid deadlock
612 Granular locking is not enough to speed up operations, we also need a
613 queue to store these and to be able to process as many as possible in
616 A Ganeti job will consist of multiple ``OpCodes`` which are the basic
617 element of operation in Ganeti 1.2 (and will remain as such). Most
618 command-level commands are equivalent to one OpCode, or in some cases
619 to a sequence of opcodes, all of the same type (e.g. evacuating a node
620 will generate N opcodes of type replace disks).
623 Job execution—“Life of a Ganeti job”
624 ++++++++++++++++++++++++++++++++++++
626 #. Job gets submitted by the client. A new job identifier is generated
627 and assigned to the job. The job is then automatically replicated
628 [#replic]_ to all nodes in the cluster. The identifier is returned to
630 #. A pool of worker threads waits for new jobs. If all are busy, the job
631 has to wait and the first worker finishing its work will grab it.
632 Otherwise any of the waiting threads will pick up the new job.
633 #. Client waits for job status updates by calling a waiting RPC
634 function. Log message may be shown to the user. Until the job is
635 started, it can also be canceled.
636 #. As soon as the job is finished, its final result and status can be
637 retrieved from the server.
638 #. If the client archives the job, it gets moved to a history directory.
639 There will be a method to archive all jobs older than a a given age.
641 .. [#replic] We need replication in order to maintain the consistency
642 across all nodes in the system; the master node only differs in the
643 fact that now it is running the master daemon, but it if fails and we
644 do a master failover, the jobs are still visible on the new master
645 (though marked as failed).
647 Failures to replicate a job to other nodes will be only flagged as
648 errors in the master daemon log if more than half of the nodes failed,
649 otherwise we ignore the failure, and rely on the fact that the next
650 update (for still running jobs) will retry the update. For finished
651 jobs, it is less of a problem.
653 Future improvements will look into checking the consistency of the job
654 list and jobs themselves at master daemon startup.
660 Jobs are stored in the filesystem as individual files, serialized
661 using JSON (standard serialization mechanism in Ganeti).
663 The choice of storing each job in its own file was made because:
665 - a file can be atomically replaced
666 - a file can easily be replicated to other nodes
667 - checking consistency across nodes can be implemented very easily,
668 since all job files should be (at a given moment in time) identical
670 The other possible choices that were discussed and discounted were:
672 - single big file with all job data: not feasible due to difficult
674 - in-process databases: hard to replicate the entire database to the
675 other nodes, and replicating individual operations does not mean wee
682 All file operations have to be done atomically by writing to a temporary
683 file and subsequent renaming. Except for log messages, every change in a
684 job is stored and replicated to other nodes.
688 /var/lib/ganeti/queue/
689 job-1 (JSON encoded job description and status)
694 lock (Queue managing process opens this file in exclusive mode)
695 serial (Last job ID used)
696 version (Queue format version)
702 Locking in the job queue is a complicated topic. It is called from more
703 than one thread and must be thread-safe. For simplicity, a single lock
704 is used for the whole job queue.
706 A more detailed description can be found in doc/locking.rst.
712 RPC calls available between Ganeti master and node daemons:
714 jobqueue_update(file_name, content)
715 Writes a file in the job queue directory.
717 Cleans the job queue directory completely, including archived job.
718 jobqueue_rename(old, new)
719 Renames a file in the job queue directory.
725 RPC between Ganeti clients and the Ganeti master daemon supports the
726 following operations:
729 Submits a list of opcodes and returns the job identifier. The
730 identifier is guaranteed to be unique during the lifetime of a
732 WaitForJobChange(job_id, fields, […], timeout)
733 This function waits until a job changes or a timeout expires. The
734 condition for when a job changed is defined by the fields passed and
735 the last log message received.
736 QueryJobs(job_ids, fields)
737 Returns field values for the job identifiers passed.
739 Cancels the job specified by identifier. This operation may fail if
740 the job is already running, canceled or finished.
742 Moves a job into the …/archive/ directory. This operation will fail if
743 the job has not been canceled or finished.
746 Job and opcode status
747 +++++++++++++++++++++
749 Each job and each opcode has, at any time, one of the following states:
752 The job/opcode was submitted, but did not yet start.
754 The job/opcode is waiting for a lock to proceed.
756 The job/opcode is running.
758 The job/opcode was canceled before it started.
760 The job/opcode ran and finished successfully.
762 The job/opcode was aborted with an error.
764 If the master is aborted while a job is running, the job will be set to
765 the Error status once the master started again.
771 Archived jobs are kept in a separate directory,
772 ``/var/lib/ganeti/queue/archive/``. This is done in order to speed up
773 the queue handling: by default, the jobs in the archive are not
774 touched by any functions. Only the current (unarchived) jobs are
775 parsed, loaded, and verified (if implemented) by the master daemon.
781 The queue has to be completely empty for Ganeti updates with changes
782 in the job queue structure. In order to allow this, there will be a
783 way to prevent new jobs entering the queue.
789 Across all cluster configuration data, we have multiple classes of
792 A. cluster-wide parameters (e.g. name of the cluster, the master);
793 these are the ones that we have today, and are unchanged from the
798 #. instance specific parameters, e.g. the name of disks (LV), that
799 cannot be shared with other instances
801 #. instance parameters, that are or can be the same for many
802 instances, but are not hypervisor related; e.g. the number of VCPUs,
803 or the size of memory
805 #. instance parameters that are hypervisor specific (e.g. kernel_path
809 The following definitions for instance parameters will be used below:
811 :hypervisor parameter:
812 a hypervisor parameter (or hypervisor specific parameter) is defined
813 as a parameter that is interpreted by the hypervisor support code in
814 Ganeti and usually is specific to a particular hypervisor (like the
815 kernel path for :term:`PVM` which makes no sense for :term:`HVM`).
818 a backend parameter is defined as an instance parameter that can be
819 shared among a list of instances, and is either generic enough not
820 to be tied to a given hypervisor or cannot influence at all the
821 hypervisor behaviour.
823 For example: memory, vcpus, auto_balance
825 All these parameters will be encoded into constants.py with the prefix
826 "BE\_" and the whole list of parameters will exist in the set
830 a parameter whose value is unique to the instance (e.g. the name of a
831 LV, or the MAC of a NIC)
833 As a general rule, for all kind of parameters, “None” (or in
834 JSON-speak, “nil”) will no longer be a valid value for a parameter. As
835 such, only non-default parameters will be saved as part of objects in
836 the serialization step, reducing the size of the serialized format.
841 Cluster parameters remain as today, attributes at the top level of the
842 Cluster object. In addition, two new attributes at this level will
843 hold defaults for the instances:
845 - hvparams, a dictionary indexed by hypervisor type, holding default
846 values for hypervisor parameters that are not defined/overridden by
847 the instances of this hypervisor type
849 - beparams, a dictionary holding (for 2.0) a single element 'default',
850 which holds the default value for backend parameters
855 Node-related parameters are very few, and we will continue using the
856 same model for these as previously (attributes on the Node object).
858 There are three new node flags, described in a separate section "node
864 As described before, the instance parameters are split in three:
865 instance proper parameters, unique to each instance, instance
866 hypervisor parameters and instance backend parameters.
868 The “hvparams” and “beparams” are kept in two dictionaries at instance
869 level. Only non-default parameters are stored (but once customized, a
870 parameter will be kept, even with the same value as the default one,
873 The names for hypervisor parameters in the instance.hvparams subtree
874 should be choosen as generic as possible, especially if specific
875 parameters could conceivably be useful for more than one hypervisor,
876 e.g. ``instance.hvparams.vnc_console_port`` instead of using both
877 ``instance.hvparams.hvm_vnc_console_port`` and
878 ``instance.hvparams.kvm_vnc_console_port``.
880 There are some special cases related to disks and NICs (for example):
881 a disk has both Ganeti-related parameters (e.g. the name of the LV)
882 and hypervisor-related parameters (how the disk is presented to/named
883 in the instance). The former parameters remain as proper-instance
884 parameters, while the latter value are migrated to the hvparams
885 structure. In 2.0, we will have only globally-per-instance such
886 hypervisor parameters, and not per-disk ones (e.g. all NICs will be
887 exported as of the same type).
889 Starting from the 1.2 list of instance parameters, here is how they
890 will be mapped to the three classes of parameters:
905 - hvm_boot_order (HV)
908 - hvm_cdrom_image_path (HV)
911 - vnc_bind_address (HV)
918 To support the new cluster parameter design, additional features will
919 be required from the hypervisor support implementations in Ganeti.
921 The hypervisor support implementation API will be extended with the
924 :PARAMETERS: class-level attribute holding the list of valid parameters
926 :CheckParamSyntax(hvparams): checks that the given parameters are
927 valid (as in the names are valid) for this hypervisor; usually just
928 comparing ``hvparams.keys()`` and ``cls.PARAMETERS``; this is a class
929 method that can be called from within master code (i.e. cmdlib) and
930 should be safe to do so
931 :ValidateParameters(hvparams): verifies the values of the provided
932 parameters against this hypervisor; this is a method that will be
933 called on the target node, from backend.py code, and as such can
934 make node-specific checks (e.g. kernel_path checking)
936 Default value application
937 +++++++++++++++++++++++++
939 The application of defaults to an instance is done in the Cluster
940 object, via two new methods as follows:
942 - ``Cluster.FillHV(instance)``, returns 'filled' hvparams dict, based on
943 instance's hvparams and cluster's ``hvparams[instance.hypervisor]``
945 - ``Cluster.FillBE(instance, be_type="default")``, which returns the
946 beparams dict, based on the instance and cluster beparams
948 The FillHV/BE transformations will be used, for example, in the
949 RpcRunner when sending an instance for activation/stop, and the sent
950 instance hvparams/beparams will have the final value (noded code doesn't
951 know about defaults).
953 LU code will need to self-call the transformation, if needed.
958 The parameter changes will have impact on the OpCodes, especially on
961 - ``OpInstanceCreate``, where the new hv and be parameters will be sent
962 as dictionaries; note that all hv and be parameters are now optional,
963 as the values can be instead taken from the cluster
964 - ``OpInstanceQuery``, where we have to be able to query these new
965 parameters; the syntax for names will be ``hvparam/$NAME`` and
966 ``beparam/$NAME`` for querying an individual parameter out of one
967 dictionary, and ``hvparams``, respectively ``beparams``, for the whole
969 - ``OpModifyInstance``, where the the modified parameters are sent as
972 Additionally, we will need new OpCodes to modify the cluster-level
973 defaults for the be/hv sets of parameters.
978 One problem that might appear is that our classification is not
979 complete or not good enough, and we'll need to change this model. As
980 the last resort, we will need to rollback and keep 1.2 style.
982 Another problem is that classification of one parameter is unclear
983 (e.g. ``network_port``, is this BE or HV?); in this case we'll take
984 the risk of having to move parameters later between classes.
989 The only security issue that we foresee is if some new parameters will
990 have sensitive value. If so, we will need to have a way to export the
991 config data while purging the sensitive value.
993 E.g. for the drbd shared secrets, we could export these with the
994 values replaced by an empty string.
999 Ganeti 2.0 adds three node flags that change the way nodes are handled
1000 within Ganeti and the related infrastructure (iallocator interaction,
1003 *master candidate* flag
1004 +++++++++++++++++++++++
1006 Ganeti 2.0 allows more scalability in operation by introducing
1007 parallelization. However, a new bottleneck is reached that is the
1008 synchronization and replication of cluster configuration to all nodes
1011 This breaks scalability as the speed of the replication decreases
1012 roughly with the size of the nodes in the cluster. The goal of the
1013 master candidate flag is to change this O(n) into O(1) with respect to
1014 job and configuration data propagation.
1016 Only nodes having this flag set (let's call this set of nodes the
1017 *candidate pool*) will have jobs and configuration data replicated.
1019 The cluster will have a new parameter (runtime changeable) called
1020 ``candidate_pool_size`` which represents the number of candidates the
1021 cluster tries to maintain (preferably automatically).
1023 This will impact the cluster operations as follows:
1025 - jobs and config data will be replicated only to a fixed set of nodes
1026 - master fail-over will only be possible to a node in the candidate pool
1027 - cluster verify needs changing to account for these two roles
1028 - external scripts will no longer have access to the configuration
1029 file (this is not recommended anyway)
1032 The caveats of this change are:
1034 - if all candidates are lost (completely), cluster configuration is
1035 lost (but it should be backed up external to the cluster anyway)
1037 - failed nodes which are candidate must be dealt with properly, so
1038 that we don't lose too many candidates at the same time; this will be
1039 reported in cluster verify
1041 - the 'all equal' concept of ganeti is no longer true
1043 - the partial distribution of config data means that all nodes will
1044 have to revert to ssconf files for master info (as in 1.2)
1048 - speed on a 100+ nodes simulated cluster is greatly enhanced, even
1049 for a simple operation; ``gnt-instance remove`` on a diskless instance
1050 remove goes from ~9seconds to ~2 seconds
1052 - node failure of non-candidates will be less impacting on the cluster
1054 The default value for the candidate pool size will be set to 10 but
1055 this can be changed at cluster creation and modified any time later.
1057 Testing on simulated big clusters with sequential and parallel jobs
1058 show that this value (10) is a sweet-spot from performance and load
1064 In order to support better the situation in which nodes are offline
1065 (e.g. for repair) without altering the cluster configuration, Ganeti
1066 needs to be told and needs to properly handle this state for nodes.
1068 This will result in simpler procedures, and less mistakes, when the
1069 amount of node failures is high on an absolute scale (either due to
1070 high failure rate or simply big clusters).
1072 Nodes having this attribute set will not be contacted for inter-node
1073 RPC calls, will not be master candidates, and will not be able to host
1074 instances as primaries.
1076 Setting this attribute on a node:
1078 - will not be allowed if the node is the master
1079 - will not be allowed if the node has primary instances
1080 - will cause the node to be demoted from the master candidate role (if
1081 it was), possibly causing another node to be promoted to that role
1083 This attribute will impact the cluster operations as follows:
1085 - querying these nodes for anything will fail instantly in the RPC
1086 library, with a specific RPC error (RpcResult.offline == True)
1088 - they will be listed in the Other section of cluster verify
1090 The code is changed in the following ways:
1092 - RPC calls were be converted to skip such nodes:
1094 - RpcRunner-instance-based RPC calls are easy to convert
1096 - static/classmethod RPC calls are harder to convert, and were left
1099 - the RPC results were unified so that this new result state (offline)
1100 can be differentiated
1102 - master voting still queries in repair nodes, as we need to ensure
1103 consistency in case the (wrong) masters have old data, and nodes have
1104 come back from repairs
1108 - some operation semantics are less clear (e.g. what to do on instance
1109 start with offline secondary?); for now, these will just fail as if
1110 the flag is not set (but faster)
1111 - 2-node cluster with one node offline needs manual startup of the
1112 master with a special flag to skip voting (as the master can't get a
1115 One of the advantages of implementing this flag is that it will allow
1116 in the future automation tools to automatically put the node in
1117 repairs and recover from this state, and the code (should/will) handle
1118 this much better than just timing out. So, future possible
1119 improvements (for later versions):
1121 - watcher will detect nodes which fail RPC calls, will attempt to ssh
1122 to them, if failure will put them offline
1123 - watcher will try to ssh and query the offline nodes, if successful
1124 will take them off the repair list
1126 Alternatives considered: The RPC call model in 2.0 is, by default,
1127 much nicer - errors are logged in the background, and job/opcode
1128 execution is clearer, so we could simply not introduce this. However,
1129 having this state will make both the codepaths clearer (offline
1130 vs. temporary failure) and the operational model (it's not a node with
1131 errors, but an offline node).
1137 Due to parallel execution of jobs in Ganeti 2.0, we could have the
1138 following situation:
1140 - gnt-node migrate + failover is run
1141 - gnt-node evacuate is run, which schedules a long-running 6-opcode
1143 - partway through, a new job comes in that runs an iallocator script,
1144 which finds the above node as empty and a very good candidate
1145 - gnt-node evacuate has finished, but now it has to be run again, to
1146 clean the above instance(s)
1148 In order to prevent this situation, and to be able to get nodes into
1149 proper offline status easily, a new *drained* flag was added to the
1152 This flag (which actually means "is being, or was drained, and is
1153 expected to go offline"), will prevent allocations on the node, but
1154 otherwise all other operations (start/stop instance, query, etc.) are
1155 working without any restrictions.
1157 Interaction between flags
1158 +++++++++++++++++++++++++
1160 While these flags are implemented as separate flags, they are
1161 mutually-exclusive and are acting together with the master node role
1162 as a single *node status* value. In other words, a flag is only in one
1163 of these roles at a given time. The lack of any of these flags denote
1166 The current node status is visible in the ``gnt-cluster verify``
1167 output, and the individual flags can be examined via separate flags in
1168 the ``gnt-node list`` output.
1170 These new flags will be exported in both the iallocator input message
1171 and via RAPI, see the respective man pages for the exact names.
1176 The main feature-level changes will be:
1178 - a number of disk related changes
1179 - removal of fixed two-disk, one-nic per instance limitation
1181 Disk handling changes
1182 ~~~~~~~~~~~~~~~~~~~~~
1184 The storage options available in Ganeti 1.x were introduced based on
1185 then-current software (first DRBD 0.7 then later DRBD 8) and the
1186 estimated usage patters. However, experience has later shown that some
1187 assumptions made initially are not true and that more flexibility is
1190 One main assumption made was that disk failures should be treated as
1191 'rare' events, and that each of them needs to be manually handled in
1192 order to ensure data safety; however, both these assumptions are false:
1194 - disk failures can be a common occurrence, based on usage patterns or
1196 - our disk setup is robust enough (referring to DRBD8 + LVM) that we
1197 could automate more of the recovery
1199 Note that we still don't have fully-automated disk recovery as a goal,
1200 but our goal is to reduce the manual work needed.
1202 As such, we plan the following main changes:
1204 - DRBD8 is much more flexible and stable than its previous version
1205 (0.7), such that removing the support for the ``remote_raid1``
1206 template and focusing only on DRBD8 is easier
1208 - dynamic discovery of DRBD devices is not actually needed in a cluster
1209 that where the DRBD namespace is controlled by Ganeti; switching to a
1210 static assignment (done at either instance creation time or change
1211 secondary time) will change the disk activation time from O(n) to
1212 O(1), which on big clusters is a significant gain
1214 - remove the hard dependency on LVM (currently all available storage
1215 types are ultimately backed by LVM volumes) by introducing file-based
1218 Additionally, a number of smaller enhancements are also planned:
1219 - support variable number of disks
1220 - support read-only disks
1222 Future enhancements in the 2.x series, which do not require base design
1223 changes, might include:
1225 - enhancement of the LVM allocation method in order to try to keep
1226 all of an instance's virtual disks on the same physical
1229 - add support for DRBD8 authentication at handshake time in
1230 order to ensure each device connects to the correct peer
1232 - remove the restrictions on failover only to the secondary
1233 which creates very strict rules on cluster allocation
1235 DRBD minor allocation
1236 +++++++++++++++++++++
1238 Currently, when trying to identify or activate a new DRBD (or MD)
1239 device, the code scans all in-use devices in order to see if we find
1240 one that looks similar to our parameters and is already in the desired
1241 state or not. Since this needs external commands to be run, it is very
1242 slow when more than a few devices are already present.
1244 Therefore, we will change the discovery model from dynamic to
1245 static. When a new device is logically created (added to the
1246 configuration) a free minor number is computed from the list of
1247 devices that should exist on that node and assigned to that
1250 At device activation, if the minor is already in use, we check if
1251 it has our parameters; if not so, we just destroy the device (if
1252 possible, otherwise we abort) and start it with our own
1255 This means that we in effect take ownership of the minor space for
1256 that device type; if there's a user-created DRBD minor, it will be
1257 automatically removed.
1259 The change will have the effect of reducing the number of external
1260 commands run per device from a constant number times the index of the
1261 first free DRBD minor to just a constant number.
1263 Removal of obsolete device types (MD, DRBD7)
1264 ++++++++++++++++++++++++++++++++++++++++++++
1266 We need to remove these device types because of two issues. First,
1267 DRBD7 has bad failure modes in case of dual failures (both network and
1268 disk - it cannot propagate the error up the device stack and instead
1269 just panics. Second, due to the asymmetry between primary and
1270 secondary in MD+DRBD mode, we cannot do live failover (not even if we
1273 File-based storage support
1274 ++++++++++++++++++++++++++
1276 Using files instead of logical volumes for instance storage would
1277 allow us to get rid of the hard requirement for volume groups for
1278 testing clusters and it would also allow usage of SAN storage to do
1279 live failover taking advantage of this storage solution.
1281 Better LVM allocation
1282 +++++++++++++++++++++
1284 Currently, the LV to PV allocation mechanism is a very simple one: at
1285 each new request for a logical volume, tell LVM to allocate the volume
1286 in order based on the amount of free space. This is good for
1287 simplicity and for keeping the usage equally spread over the available
1288 physical disks, however it introduces a problem that an instance could
1289 end up with its (currently) two drives on two physical disks, or
1290 (worse) that the data and metadata for a DRBD device end up on
1293 This is bad because it causes unneeded ``replace-disks`` operations in
1294 case of a physical failure.
1296 The solution is to batch allocations for an instance and make the LVM
1297 handling code try to allocate as close as possible all the storage of
1298 one instance. We will still allow the logical volumes to spill over to
1299 additional disks as needed.
1301 Note that this clustered allocation can only be attempted at initial
1302 instance creation, or at change secondary node time. At add disk time,
1303 or at replacing individual disks, it's not easy enough to compute the
1304 current disk map so we'll not attempt the clustering.
1306 DRBD8 peer authentication at handshake
1307 ++++++++++++++++++++++++++++++++++++++
1309 DRBD8 has a new feature that allow authentication of the peer at
1310 connect time. We can use this to prevent connecting to the wrong peer
1311 more that securing the connection. Even though we never had issues
1312 with wrong connections, it would be good to implement this.
1315 LVM self-repair (optional)
1316 ++++++++++++++++++++++++++
1318 The complete failure of a physical disk is very tedious to
1319 troubleshoot, mainly because of the many failure modes and the many
1320 steps needed. We can safely automate some of the steps, more
1321 specifically the ``vgreduce --removemissing`` using the following
1324 #. check if all nodes have consistent volume groups
1325 #. if yes, and previous status was yes, do nothing
1326 #. if yes, and previous status was no, save status and restart
1327 #. if no, and previous status was no, do nothing
1328 #. if no, and previous status was yes:
1329 #. if more than one node is inconsistent, do nothing
1330 #. if only one node is inconsistent:
1331 #. run ``vgreduce --removemissing``
1332 #. log this occurrence in the Ganeti log in a form that
1333 can be used for monitoring
1334 #. [FUTURE] run ``replace-disks`` for all
1337 Failover to any node
1338 ++++++++++++++++++++
1340 With a modified disk activation sequence, we can implement the
1341 *failover to any* functionality, removing many of the layout
1342 restrictions of a cluster:
1344 - the need to reserve memory on the current secondary: this gets reduced
1345 to a must to reserve memory anywhere on the cluster
1347 - the need to first failover and then replace secondary for an
1348 instance: with failover-to-any, we can directly failover to
1349 another node, which also does the replace disks at the same
1352 In the following, we denote the current primary by P1, the current
1353 secondary by S1, and the new primary and secondaries by P2 and S2. P2
1354 is fixed to the node the user chooses, but the choice of S2 can be
1355 made between P1 and S1. This choice can be constrained, depending on
1356 which of P1 and S1 has failed.
1358 - if P1 has failed, then S1 must become S2, and live migration is not
1360 - if S1 has failed, then P1 must become S2, and live migration could be
1361 possible (in theory, but this is not a design goal for 2.0)
1363 The algorithm for performing the failover is straightforward:
1365 - verify that S2 (the node the user has chosen to keep as secondary) has
1366 valid data (is consistent)
1368 - tear down the current DRBD association and setup a DRBD pairing
1369 between P2 (P2 is indicated by the user) and S2; since P2 has no data,
1370 it will start re-syncing from S2
1372 - as soon as P2 is in state SyncTarget (i.e. after the resync has
1373 started but before it has finished), we can promote it to primary role
1374 (r/w) and start the instance on P2
1376 - as soon as the P2?S2 sync has finished, we can remove
1377 the old data on the old node that has not been chosen for
1380 Caveats: during the P2?S2 sync, a (non-transient) network error
1381 will cause I/O errors on the instance, so (if a longer instance
1382 downtime is acceptable) we can postpone the restart of the instance
1383 until the resync is done. However, disk I/O errors on S2 will cause
1384 data loss, since we don't have a good copy of the data anymore, so in
1385 this case waiting for the sync to complete is not an option. As such,
1386 it is recommended that this feature is used only in conjunction with
1387 proper disk monitoring.
1390 Live migration note: While failover-to-any is possible for all choices
1391 of S2, migration-to-any is possible only if we keep P1 as S2.
1396 The dynamic device model, while more complex, has an advantage: it
1397 will not reuse by mistake the DRBD device of another instance, since
1398 it always looks for either our own or a free one.
1400 The static one, in contrast, will assume that given a minor number N,
1401 it's ours and we can take over. This needs careful implementation such
1402 that if the minor is in use, either we are able to cleanly shut it
1403 down, or we abort the startup. Otherwise, it could be that we start
1404 syncing between two instance's disks, causing data loss.
1407 Variable number of disk/NICs per instance
1408 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1410 Variable number of disks
1411 ++++++++++++++++++++++++
1413 In order to support high-security scenarios (for example read-only sda
1414 and read-write sdb), we need to make a fully flexibly disk
1415 definition. This has less impact that it might look at first sight:
1416 only the instance creation has hard coded number of disks, not the disk
1417 handling code. The block device handling and most of the instance
1418 handling code is already working with "the instance's disks" as
1419 opposed to "the two disks of the instance", but some pieces are not
1420 (e.g. import/export) and the code needs a review to ensure safety.
1422 The objective is to be able to specify the number of disks at
1423 instance creation, and to be able to toggle from read-only to
1424 read-write a disk afterward.
1426 Variable number of NICs
1427 +++++++++++++++++++++++
1429 Similar to the disk change, we need to allow multiple network
1430 interfaces per instance. This will affect the internal code (some
1431 function will have to stop assuming that ``instance.nics`` is a list
1432 of length one), the OS API which currently can export/import only one
1433 instance, and the command line interface.
1438 There are two areas of interface changes: API-level changes (the OS
1439 interface and the RAPI interface) and the command line interface
1445 The current Ganeti OS interface, version 5, is tailored for Ganeti 1.2.
1446 The interface is composed by a series of scripts which get called with
1447 certain parameters to perform OS-dependent operations on the cluster.
1448 The current scripts are:
1451 called when a new instance is added to the cluster
1453 called to export an instance disk to a stream
1455 called to import from a stream to a new instance
1457 called to perform the os-specific operations necessary for renaming an
1460 Currently these scripts suffer from the limitations of Ganeti 1.2: for
1461 example they accept exactly one block and one swap devices to operate
1462 on, rather than any amount of generic block devices, they blindly assume
1463 that an instance will have just one network interface to operate, they
1464 can not be configured to optimise the instance for a particular
1467 Since in Ganeti 2.0 we want to support multiple hypervisors, and a
1468 non-fixed number of network and disks the OS interface need to change to
1469 transmit the appropriate amount of information about an instance to its
1470 managing operating system, when operating on it. Moreover since some old
1471 assumptions usually used in OS scripts are no longer valid we need to
1472 re-establish a common knowledge on what can be assumed and what cannot
1473 be regarding Ganeti environment.
1476 When designing the new OS API our priorities are:
1478 - future extensibility
1479 - ease of porting from the old API
1482 As such we want to limit the number of scripts that must be written to
1483 support an OS, and make it easy to share code between them by uniforming
1484 their input. We also will leave the current script structure unchanged,
1485 as far as we can, and make a few of the scripts (import, export and
1486 rename) optional. Most information will be passed to the script through
1487 environment variables, for ease of access and at the same time ease of
1488 using only the information a script needs.
1494 As in Ganeti 1.2, every OS which wants to be installed in Ganeti needs
1495 to support the following functionality, through scripts:
1498 used to create a new instance running that OS. This script should
1499 prepare the block devices, and install them so that the new OS can
1500 boot under the specified hypervisor.
1502 used to export an installed instance using the given OS to a format
1503 which can be used to import it back into a new instance.
1505 used to import an exported instance into a new one. This script is
1506 similar to create, but the new instance should have the content of the
1507 export, rather than contain a pristine installation.
1509 used to perform the internal OS-specific operations needed to rename
1512 If any optional script is not implemented Ganeti will refuse to perform
1513 the given operation on instances using the non-implementing OS. Of
1514 course the create script is mandatory, and it doesn't make sense to
1515 support the either the export or the import operation but not both.
1517 Incompatibilities with 1.2
1518 __________________________
1520 We expect the following incompatibilities between the OS scripts for 1.2
1521 and the ones for 2.0:
1523 - Input parameters: in 1.2 those were passed on the command line, in 2.0
1524 we'll use environment variables, as there will be a lot more
1525 information and not all OSes may care about all of it.
1526 - Number of calls: export scripts will be called once for each device
1527 the instance has, and import scripts once for every exported disk.
1528 Imported instances will be forced to have a number of disks greater or
1529 equal to the one of the export.
1530 - Some scripts are not compulsory: if such a script is missing the
1531 relevant operations will be forbidden for instances of that OS. This
1532 makes it easier to distinguish between unsupported operations and
1533 no-op ones (if any).
1539 Rather than using command line flags, as they do now, scripts will
1540 accept inputs from environment variables. We expect the following input
1544 The version of the OS API that the following parameters comply with;
1545 this is used so that in the future we could have OSes supporting
1546 multiple versions and thus Ganeti send the proper version in this
1549 Name of the instance acted on
1551 The hypervisor the instance should run on (e.g. 'xen-pvm', 'xen-hvm',
1554 The number of disks this instance will have
1556 The number of NICs this instance will have
1558 Path to the Nth disk.
1560 W if read/write, R if read only. OS scripts are not supposed to touch
1561 read-only disks, but will be passed them to know.
1562 DISK_<N>_FRONTEND_TYPE
1563 Type of the disk as seen by the instance. Can be 'scsi', 'ide',
1565 DISK_<N>_BACKEND_TYPE
1566 Type of the disk as seen from the node. Can be 'block', 'file:loop' or
1569 Mac address for the Nth network interface
1571 Ip address for the Nth network interface, if available
1573 Node bridge the Nth network interface will be connected to
1574 NIC_<N>_FRONTEND_TYPE
1575 Type of the Nth NIC as seen by the instance. For example 'virtio',
1578 Whether more out should be produced, for debugging purposes. Currently
1579 the only valid values are 0 and 1.
1581 These are only the basic variables we are thinking of now, but more
1582 may come during the implementation and they will be documented in the
1583 :manpage:`ganeti-os-api` man page. All these variables will be
1584 available to all scripts.
1586 Some scripts will need a few more information to work. These will have
1587 per-script variables, such as for example:
1590 rename: the name the instance should be renamed from.
1592 export: device to be exported, a snapshot of the actual device. The
1593 data must be exported to stdout.
1595 export: sequential number of the instance device targeted.
1597 import: device to send the data to, part of the new instance. The data
1598 must be imported from stdin.
1600 import: sequential number of the instance device targeted.
1602 (Rationale for INSTANCE_NAME as an environment variable: the instance
1603 name is always needed and we could pass it on the command line. On the
1604 other hand, though, this would force scripts to both access the
1605 environment and parse the command line, so we'll move it for
1612 As discussed scripts should only send user-targeted information to
1613 stderr. The create and import scripts are supposed to format/initialise
1614 the given block devices and install the correct instance data. The
1615 export script is supposed to export instance data to stdout in a format
1616 understandable by the the import script. The data will be compressed by
1617 Ganeti, so no compression should be done. The rename script should only
1618 modify the instance's knowledge of what its name is.
1620 Other declarative style features
1621 ++++++++++++++++++++++++++++++++
1623 Similar to Ganeti 1.2, OS specifications will need to provide a
1624 'ganeti_api_version' containing list of numbers matching the
1625 version(s) of the API they implement. Ganeti itself will always be
1626 compatible with one version of the API and may maintain backwards
1627 compatibility if it's feasible to do so. The numbers are one-per-line,
1628 so an OS supporting both version 5 and version 20 will have a file
1629 containing two lines. This is different from Ganeti 1.2, which only
1630 supported one version number.
1632 In addition to that an OS will be able to declare that it does support
1633 only a subset of the Ganeti hypervisors, by declaring them in the
1640 We might want to have a "default" import/export behaviour that just
1641 dumps all disks and restores them. This can save work as most systems
1642 will just do this, while allowing flexibility for different systems.
1644 Environment variables are limited in size, but we expect that there will
1645 be enough space to store the information we need. If we discover that
1646 this is not the case we may want to go to a more complex API such as
1647 storing those information on the filesystem and providing the OS script
1648 with the path to a file where they are encoded in some format.
1655 The first Ganeti remote API (RAPI) was designed and deployed with the
1656 Ganeti 1.2.5 release. That version provide read-only access to the
1657 cluster state. Fully functional read-write API demands significant
1658 internal changes which will be implemented in version 2.0.
1660 We decided to go with implementing the Ganeti RAPI in a RESTful way,
1661 which is aligned with key features we looking. It is simple,
1662 stateless, scalable and extensible paradigm of API implementation. As
1663 transport it uses HTTP over SSL, and we are implementing it with JSON
1664 encoding, but in a way it possible to extend and provide any other
1670 The Ganeti RAPI is implemented as independent daemon, running on the
1671 same node with the same permission level as Ganeti master
1672 daemon. Communication is done through the LUXI library to the master
1673 daemon. In order to keep communication asynchronous RAPI processes two
1674 types of client requests:
1676 - queries: server is able to answer immediately
1677 - job submission: some time is required for a useful response
1679 In the query case requested data send back to client in the HTTP
1680 response body. Typical examples of queries would be: list of nodes,
1681 instances, cluster info, etc.
1683 In the case of job submission, the client receive a job ID, the
1684 identifier which allows to query the job progress in the job queue
1687 Internally, each exported object has an version identifier, which is
1688 used as a state identifier in the HTTP header E-Tag field for
1689 requests/responses to avoid race conditions.
1692 Resource representation
1693 +++++++++++++++++++++++
1695 The key difference of using REST instead of others API is that REST
1696 requires separation of services via resources with unique URIs. Each
1697 of them should have limited amount of state and support standard HTTP
1698 methods: GET, POST, DELETE, PUT.
1700 For example in Ganeti's case we can have a set of URI:
1702 - ``/{clustername}/instances``
1703 - ``/{clustername}/instances/{instancename}``
1704 - ``/{clustername}/instances/{instancename}/tag``
1705 - ``/{clustername}/tag``
1707 A GET request to ``/{clustername}/instances`` will return the list of
1708 instances, a POST to ``/{clustername}/instances`` should create a new
1709 instance, a DELETE ``/{clustername}/instances/{instancename}`` should
1710 delete the instance, a GET ``/{clustername}/tag`` should return get
1713 Each resource URI will have a version prefix. The resource IDs are to
1716 Internal encoding might be JSON, XML, or any other. The JSON encoding
1717 fits nicely in Ganeti RAPI needs. The client can request a specific
1718 representation via the Accept field in the HTTP header.
1720 REST uses HTTP as its transport and application protocol for resource
1721 access. The set of possible responses is a subset of standard HTTP
1724 The statelessness model provides additional reliability and
1725 transparency to operations (e.g. only one request needs to be analyzed
1726 to understand the in-progress operation, not a sequence of multiple
1727 requests/responses).
1733 With the write functionality security becomes a much bigger an issue.
1734 The Ganeti RAPI uses basic HTTP authentication on top of an
1735 SSL-secured connection to grant access to an exported resource. The
1736 password is stored locally in an Apache-style ``.htpasswd`` file. Only
1737 one level of privileges is supported.
1742 The model detailed above for job submission requires the client to
1743 poll periodically for updates to the job; an alternative would be to
1744 allow the client to request a callback, or a 'wait for updates' call.
1746 The callback model was not considered due to the following two issues:
1748 - callbacks would require a new model of allowed callback URLs,
1749 together with a method of managing these
1750 - callbacks only work when the client and the master are in the same
1751 security domain, and they fail in the other cases (e.g. when there is
1752 a firewall between the client and the RAPI daemon that only allows
1753 client-to-RAPI calls, which is usual in DMZ cases)
1755 The 'wait for updates' method is not suited to the HTTP protocol,
1756 where requests are supposed to be short-lived.
1758 Command line changes
1759 ~~~~~~~~~~~~~~~~~~~~
1761 Ganeti 2.0 introduces several new features as well as new ways to
1762 handle instance resources like disks or network interfaces. This
1763 requires some noticeable changes in the way command line arguments are
1766 - extend and modify command line syntax to support new features
1767 - ensure consistent patterns in command line arguments to reduce
1770 The design changes that require these changes are, in no particular
1773 - flexible instance disk handling: support a variable number of disks
1774 with varying properties per instance,
1775 - flexible instance network interface handling: support a variable
1776 number of network interfaces with varying properties per instance
1777 - multiple hypervisors: multiple hypervisors can be active on the same
1778 cluster, each supporting different parameters,
1779 - support for device type CDROM (via ISO image)
1781 As such, there are several areas of Ganeti where the command line
1782 arguments will change:
1784 - Cluster configuration
1786 - cluster initialization
1787 - cluster default configuration
1789 - Instance configuration
1791 - handling of network cards for instances,
1792 - handling of disks for instances,
1793 - handling of CDROM devices and
1794 - handling of hypervisor specific options.
1796 There are several areas of Ganeti where the command line arguments
1799 - Cluster configuration
1801 - cluster initialization
1802 - cluster default configuration
1804 - Instance configuration
1806 - handling of network cards for instances,
1807 - handling of disks for instances,
1808 - handling of CDROM devices and
1809 - handling of hypervisor specific options.
1811 Notes about device removal/addition
1812 +++++++++++++++++++++++++++++++++++
1814 To avoid problems with device location changes (e.g. second network
1815 interface of the instance becoming the first or third and the like)
1816 the list of network/disk devices is treated as a stack, i.e. devices
1817 can only be added/removed at the end of the list of devices of each
1818 class (disk or network) for each instance.
1820 gnt-instance commands
1821 +++++++++++++++++++++
1823 The commands for gnt-instance will be modified and extended to allow
1824 for the new functionality:
1826 - the add command will be extended to support the new device and
1828 - the modify command continues to handle all modifications to
1829 instances, but will be extended with new arguments for handling
1832 Network Device Options
1833 ++++++++++++++++++++++
1835 The generic format of the network device option is:
1837 --net $DEVNUM[:$OPTION=$VALUE][,$OPTION=VALUE]
1839 :$DEVNUM: device number, unsigned integer, starting at 0,
1840 :$OPTION: device option, string,
1841 :$VALUE: device option value, string.
1843 Currently, the following device options will be defined (open to
1846 :mac: MAC address of the network interface, accepts either a valid
1847 MAC address or the string 'auto'. If 'auto' is specified, a new MAC
1848 address will be generated randomly. If the mac device option is not
1849 specified, the default value 'auto' is assumed.
1850 :bridge: network bridge the network interface is connected
1851 to. Accepts either a valid bridge name (the specified bridge must
1852 exist on the node(s)) as string or the string 'auto'. If 'auto' is
1853 specified, the default brigde is used. If the bridge option is not
1854 specified, the default value 'auto' is assumed.
1859 The generic format of the disk device option is:
1861 --disk $DEVNUM[:$OPTION=$VALUE][,$OPTION=VALUE]
1863 :$DEVNUM: device number, unsigned integer, starting at 0,
1864 :$OPTION: device option, string,
1865 :$VALUE: device option value, string.
1867 Currently, the following device options will be defined (open to
1870 :size: size of the disk device, either a positive number, specifying
1871 the disk size in mebibytes, or a number followed by a magnitude suffix
1872 (M for mebibytes, G for gibibytes). Also accepts the string 'auto' in
1873 which case the default disk size will be used. If the size option is
1874 not specified, 'auto' is assumed. This option is not valid for all
1876 :access: access mode of the disk device, a single letter, valid values
1879 - *w*: read/write access to the disk device or
1880 - *r*: read-only access to the disk device.
1882 If the access mode is not specified, the default mode of read/write
1883 access will be configured.
1884 :path: path to the image file for the disk device, string. No default
1885 exists. This option is not valid for all disk layout types.
1890 To add devices to an already existing instance, use the device type
1891 specific option to gnt-instance modify. Currently, there are two
1892 device type specific options supported:
1894 :--net: for network interface cards
1895 :--disk: for disk devices
1897 The syntax to the device specific options is similar to the generic
1898 device options, but instead of specifying a device number like for
1899 gnt-instance add, you specify the magic string add. The new device
1900 will always be appended at the end of the list of devices of this type
1901 for the specified instance, e.g. if the instance has disk devices 0,1
1902 and 2, the newly added disk device will be disk device 3.
1904 Example: gnt-instance modify --net add:mac=auto test-instance
1909 Removing devices from and instance is done via gnt-instance
1910 modify. The same device specific options as for adding instances are
1911 used. Instead of a device number and further device options, only the
1912 magic string remove is specified. It will always remove the last
1913 device in the list of devices of this type for the instance specified,
1914 e.g. if the instance has disk devices 0, 1, 2 and 3, the disk device
1915 number 3 will be removed.
1917 Example: gnt-instance modify --net remove test-instance
1922 Modifying devices is also done with device type specific options to
1923 the gnt-instance modify command. There are currently two device type
1926 :--net: for network interface cards
1927 :--disk: for disk devices
1929 The syntax to the device specific options is similar to the generic
1930 device options. The device number you specify identifies the device to
1935 gnt-instance modify --disk 2:access=r
1940 Ganeti 2.0 will support more than one hypervisor. Different
1941 hypervisors have various options that only apply to a specific
1942 hypervisor. Those hypervisor specific options are treated specially
1943 via the ``--hypervisor`` option. The generic syntax of the hypervisor
1944 option is as follows::
1946 --hypervisor $HYPERVISOR:$OPTION=$VALUE[,$OPTION=$VALUE]
1948 :$HYPERVISOR: symbolic name of the hypervisor to use, string,
1949 has to match the supported hypervisors. Example: xen-pvm
1951 :$OPTION: hypervisor option name, string
1952 :$VALUE: hypervisor option value, string
1954 The hypervisor option for an instance can be set on instance creation
1955 time via the ``gnt-instance add`` command. If the hypervisor for an
1956 instance is not specified upon instance creation, the default
1957 hypervisor will be used.
1959 Modifying hypervisor parameters
1960 +++++++++++++++++++++++++++++++
1962 The hypervisor parameters of an existing instance can be modified
1963 using ``--hypervisor`` option of the ``gnt-instance modify``
1964 command. However, the hypervisor type of an existing instance can not
1965 be changed, only the particular hypervisor specific option can be
1966 changed. Therefore, the format of the option parameters has been
1967 simplified to omit the hypervisor name and only contain the comma
1968 separated list of option-value pairs.
1972 gnt-instance modify --hypervisor cdrom=/srv/boot.iso,boot_order=cdrom:network test-instance
1974 gnt-cluster commands
1975 ++++++++++++++++++++
1977 The command for gnt-cluster will be extended to allow setting and
1978 changing the default parameters of the cluster:
1980 - The init command will be extend to support the defaults option to
1981 set the cluster defaults upon cluster initialization.
1982 - The modify command will be added to modify the cluster
1983 parameters. It will support the --defaults option to change the
1988 The generic format of the cluster default setting option is:
1990 --defaults $OPTION=$VALUE[,$OPTION=$VALUE]
1992 :$OPTION: cluster default option, string,
1993 :$VALUE: cluster default option value, string.
1995 Currently, the following cluster default options are defined (open to
1998 :hypervisor: the default hypervisor to use for new instances,
1999 string. Must be a valid hypervisor known to and supported by the
2001 :disksize: the disksize for newly created instance disks, where
2002 applicable. Must be either a positive number, in which case the unit
2003 of megabyte is assumed, or a positive number followed by a supported
2004 magnitude symbol (M for megabyte or G for gigabyte).
2005 :bridge: the default network bridge to use for newly created instance
2006 network interfaces, string. Must be a valid bridge name of a bridge
2007 existing on the node(s).
2009 Hypervisor cluster defaults
2010 +++++++++++++++++++++++++++
2012 The generic format of the hypervisor cluster wide default setting
2015 --hypervisor-defaults $HYPERVISOR:$OPTION=$VALUE[,$OPTION=$VALUE]
2017 :$HYPERVISOR: symbolic name of the hypervisor whose defaults you want
2019 :$OPTION: cluster default option, string,
2020 :$VALUE: cluster default option value, string.
2022 .. vim: set textwidth=72 :