Synnefo » snf-ganeti

=================

Ganeti 2.0 design

=================

This document describes the major changes in Ganeti 2.0 compared to

the 1.2 version.

The 2.0 version will constitute a rewrite of the 'core' architecture,

paving the way for additional features in future 2.x versions.

.. contents:: :depth: 3

Objective

=========

Ganeti 1.2 has many scalability issues and restrictions due to its

roots as software for managing small and 'static' clusters.

Version 2.0 will attempt to remedy first the scalability issues and

then the restrictions.

Background

==========

While Ganeti 1.2 is usable, it severely limits the flexibility of the

cluster administration and imposes a very rigid model. It has the

following main scalability issues:

- only one operation at a time on the cluster [#]_

- poor handling of node failures in the cluster

- mixing hypervisors in a cluster not allowed

It also has a number of artificial restrictions, due to historical

design:

- fixed number of disks (two) per instance

- fixed number of NICs

.. [#] Replace disks will release the lock, but this is an exception

       and not a recommended way to operate

The 2.0 version is intended to address some of these problems, and

create a more flexible code base for future developments.

Among these problems, the single-operation at a time restriction is

biggest issue with the current version of Ganeti. It is such a big

impediment in operating bigger clusters that many times one is tempted

to remove the lock just to do a simple operation like start instance

while an OS installation is running.

Scalability problems

--------------------

Ganeti 1.2 has a single global lock, which is used for all cluster

operations.  This has been painful at various times, for example:

- It is impossible for two people to efficiently interact with a cluster

  (for example for debugging) at the same time.

- When batch jobs are running it's impossible to do other work (for

  example failovers/fixes) on a cluster.

This poses scalability problems: as clusters grow in node and instance

size it's a lot more likely that operations which one could conceive

should run in parallel (for example because they happen on different

nodes) are actually stalling each other while waiting for the global

lock, without a real reason for that to happen.

One of the main causes of this global lock (beside the higher

difficulty of ensuring data consistency in a more granular lock model)

is the fact that currently there is no long-lived process in Ganeti

that can coordinate multiple operations. Each command tries to acquire

the so called *cmd* lock and when it succeeds, it takes complete

ownership of the cluster configuration and state.

Other scalability problems are due the design of the DRBD device

model, which assumed at its creation a low (one to four) number of

instances per node, which is no longer true with today's hardware.

Artificial restrictions

-----------------------

Ganeti 1.2 (and previous versions) have a fixed two-disks, one-NIC per

instance model. This is a purely artificial restrictions, but it

touches multiple areas (configuration, import/export, command line)

that it's more fitted to a major release than a minor one.

Architecture issues

-------------------

The fact that each command is a separate process that reads the

cluster state, executes the command, and saves the new state is also

an issue on big clusters where the configuration data for the cluster

begins to be non-trivial in size.

Overview

========

In order to solve the scalability problems, a rewrite of the core

design of Ganeti is required. While the cluster operations themselves

won't change (e.g. start instance will do the same things, the way

these operations are scheduled internally will change radically.

The new design will change the cluster architecture to:

.. digraph:: "ganeti-2.0-architecture"

  compound=false

  concentrate=true

  mclimit=100.0

  nslimit=100.0

  edge[fontsize="8" fontname="Helvetica-Oblique"]

  node[width="0" height="0" fontsize="12" fontcolor="black" shape=rect]

  subgraph outside {

    rclient[label="external clients"]

    label="Outside the cluster"

  }

  subgraph cluster_inside {

    label="ganeti cluster"

    labeljust=l

    subgraph cluster_master_node {

      label="master node"

      rapi[label="RAPI daemon"]

      cli[label="CLI"]

      watcher[label="Watcher"]

      burnin[label="Burnin"]

      masterd[shape=record style=filled label="{ <luxi> luxi endpoint | master I/O thread | job queue | {<w1> worker| <w2> worker | <w3> worker }}"]

      {rapi;cli;watcher;burnin} -> masterd:luxi [label="LUXI" labelpos=100]

    }

    subgraph cluster_nodes {

        label="nodes"

        noded1 [shape=record label="{ RPC listener | Disk management | Network management | Hypervisor } "]

        noded2 [shape=record label="{ RPC listener | Disk management | Network management | Hypervisor } "]

        noded3 [shape=record label="{ RPC listener | Disk management | Network management | Hypervisor } "]

    }

    masterd:w2 -> {noded1;noded2;noded3} [label="node RPC"]

    cli -> {noded1;noded2;noded3} [label="SSH"]

  }

  rclient -> rapi [label="RAPI protocol"]

This differs from the 1.2 architecture by the addition of the master

daemon, which will be the only entity to talk to the node daemons.

Detailed design

===============

The changes for 2.0 can be split into roughly three areas:

- core changes that affect the design of the software

- features (or restriction removals) but which do not have a wide

  impact on the design

- user-level and API-level changes which translate into differences for

  the operation of the cluster

Core changes

------------

The main changes will be switching from a per-process model to a

daemon based model, where the individual gnt-* commands will be

clients that talk to this daemon (see `Master daemon`_). This will

allow us to get rid of the global cluster lock for most operations,

having instead a per-object lock (see `Granular locking`_). Also, the

daemon will be able to queue jobs, and this will allow the individual

clients to submit jobs without waiting for them to finish, and also

see the result of old requests (see `Job Queue`_).

Beside these major changes, another 'core' change but that will not be

as visible to the users will be changing the model of object attribute

storage, and separate that into name spaces (such that an Xen PVM

instance will not have the Xen HVM parameters). This will allow future

flexibility in defining additional parameters. For more details see

`Object parameters`_.

The various changes brought in by the master daemon model and the

read-write RAPI will require changes to the cluster security; we move

away from Twisted and use HTTP(s) for intra- and extra-cluster

communications. For more details, see the security document in the

doc/ directory.

Master daemon

~~~~~~~~~~~~~

In Ganeti 2.0, we will have the following *entities*:

- the master daemon (on the master node)

- the node daemon (on all nodes)

- the command line tools (on the master node)

- the RAPI daemon (on the master node)

The master-daemon related interaction paths are:

- (CLI tools/RAPI daemon) and the master daemon, via the so called

  *LUXI* API

- the master daemon and the node daemons, via the node RPC

There are also some additional interaction paths for exceptional cases:

- CLI tools might access via SSH the nodes (for ``gnt-cluster copyfile``

  and ``gnt-cluster command``)

- master failover is a special case when a non-master node will SSH

  and do node-RPC calls to the current master

The protocol between the master daemon and the node daemons will be

changed from (Ganeti 1.2) Twisted PB (perspective broker) to HTTP(S),

using a simple PUT/GET of JSON-encoded messages. This is done due to

difficulties in working with the Twisted framework and its protocols

in a multithreaded environment, which we can overcome by using a

simpler stack (see the caveats section).

The protocol between the CLI/RAPI and the master daemon will be a

custom one (called *LUXI*): on a UNIX socket on the master node, with

rights restricted by filesystem permissions, the CLI/RAPI will talk to

the master daemon using JSON-encoded messages.

The operations supported over this internal protocol will be encoded

via a python library that will expose a simple API for its

users. Internally, the protocol will simply encode all objects in JSON

format and decode them on the receiver side.

For more details about the RAPI daemon see `Remote API changes`_, and

for the node daemon see `Node daemon changes`_.

.. _luxi:

The LUXI protocol

+++++++++++++++++

As described above, the protocol for making requests or queries to the

master daemon will be a UNIX-socket based simple RPC of JSON-encoded

messages.

The choice of UNIX was in order to get rid of the need of

authentication and authorisation inside Ganeti; for 2.0, the

permissions on the Unix socket itself will determine the access

rights.

We will have two main classes of operations over this API:

- cluster query functions

- job related functions

The cluster query functions are usually short-duration, and are the

equivalent of the ``OP_QUERY_*`` opcodes in Ganeti 1.2 (and they are

internally implemented still with these opcodes). The clients are

guaranteed to receive the response in a reasonable time via a timeout.

The job-related functions will be:

- submit job

- query job (which could also be categorized in the query-functions)

- archive job (see the job queue design doc)

- wait for job change, which allows a client to wait without polling

For more details of the actual operation list, see the `Job Queue`_.

Both requests and responses will consist of a JSON-encoded message

followed by the ``ETX`` character (ASCII decimal 3), which is not a

valid character in JSON messages and thus can serve as a message

delimiter. The contents of the messages will be a dictionary with two

fields:

:method:

  the name of the method called

:args:

  the arguments to the method, as a list (no keyword arguments allowed)

Responses will follow the same format, with the two fields being:

:success:

  a boolean denoting the success of the operation

:result:

  the actual result, or error message in case of failure

There are two special value for the result field:

- in the case that the operation failed, and this field is a list of

  length two, the client library will try to interpret is as an

  exception, the first element being the exception type and the second

  one the actual exception arguments; this will allow a simple method of

  passing Ganeti-related exception across the interface

- for the *WaitForChange* call (that waits on the server for a job to

  change status), if the result is equal to ``nochange`` instead of the

  usual result for this call (a list of changes), then the library will

  internally retry the call; this is done in order to differentiate

  internally between master daemon hung and job simply not changed

Users of the API that don't use the provided python library should

take care of the above two cases.

Master daemon implementation

++++++++++++++++++++++++++++

The daemon will be based around a main I/O thread that will wait for

new requests from the clients, and that does the setup/shutdown of the

other thread (pools).

There will two other classes of threads in the daemon:

- job processing threads, part of a thread pool, and which are

  long-lived, started at daemon startup and terminated only at shutdown

  time

- client I/O threads, which are the ones that talk the local protocol

  (LUXI) to the clients, and are short-lived

Master startup/failover

+++++++++++++++++++++++

In Ganeti 1.x there is no protection against failing over the master

to a node with stale configuration. In effect, the responsibility of

correct failovers falls on the admin. This is true both for the new

master and for when an old, offline master startup.

Since in 2.x we are extending the cluster state to cover the job queue

and have a daemon that will execute by itself the job queue, we want

to have more resilience for the master role.

The following algorithm will happen whenever a node is ready to

transition to the master role, either at startup time or at node

failover:

#. read the configuration file and parse the node list

   contained within

#. query all the nodes and make sure we obtain an agreement via

   a quorum of at least half plus one nodes for the following:

    - we have the latest configuration and job list (as

      determined by the serial number on the configuration and

      highest job ID on the job queue)

    - if we are not failing over (but just starting), the

      quorum agrees that we are the designated master

    - if any of the above is false, we prevent the current operation

      (i.e. we don't become the master)

#. at this point, the node transitions to the master role

#. for all the in-progress jobs, mark them as failed, with

   reason unknown or something similar (master failed, etc.)

Since due to exceptional conditions we could have a situation in which

no node can become the master due to inconsistent data, we will have

an override switch for the master daemon startup that will assume the

current node has the right data and will replicate all the

configuration files to the other nodes.

**Note**: the above algorithm is by no means an election algorithm; it

is a *confirmation* of the master role currently held by a node.

Logging

+++++++

The logging system will be switched completely to the standard python

logging module; currently it's logging-based, but exposes a different

API, which is just overhead. As such, the code will be switched over

to standard logging calls, and only the setup will be custom.

With this change, we will remove the separate debug/info/error logs,

and instead have always one logfile per daemon model:

- master-daemon.log for the master daemon

- node-daemon.log for the node daemon (this is the same as in 1.2)

- rapi-daemon.log for the RAPI daemon logs

- rapi-access.log, an additional log file for the RAPI that will be

  in the standard HTTP log format for possible parsing by other tools

Since the :term:`watcher` will only submit jobs to the master for

startup of the instances, its log file will contain less information

than before, mainly that it will start the instance, but not the

results.

Node daemon changes

+++++++++++++++++++

The only change to the node daemon is that, since we need better

concurrency, we don't process the inter-node RPC calls in the node

daemon itself, but we fork and process each request in a separate

child.

Since we don't have many calls, and we only fork (not exec), the

overhead should be minimal.

Caveats

+++++++

A discussed alternative is to keep the current individual processes

touching the cluster configuration model. The reasons we have not

chosen this approach is:

- the speed of reading and unserializing the cluster state

  today is not small enough that we can ignore it; the addition of

  the job queue will make the startup cost even higher. While this

  runtime cost is low, it can be on the order of a few seconds on

  bigger clusters, which for very quick commands is comparable to

  the actual duration of the computation itself

- individual commands would make it harder to implement a

  fire-and-forget job request, along the lines "start this

  instance but do not wait for it to finish"; it would require a

  model of backgrounding the operation and other things that are

  much better served by a daemon-based model

Another area of discussion is moving away from Twisted in this new

implementation. While Twisted has its advantages, there are also many

disadvantages to using it:

- first and foremost, it's not a library, but a framework; thus, if

  you use twisted, all the code needs to be 'twiste-ized' and written

  in an asynchronous manner, using deferreds; while this method works,

  it's not a common way to code and it requires that the entire process

  workflow is based around a single *reactor* (Twisted name for a main

  loop)

- the more advanced granular locking that we want to implement would

  require, if written in the async-manner, deep integration with the

  Twisted stack, to such an extend that business-logic is inseparable

  from the protocol coding; we felt that this is an unreasonable

  request, and that a good protocol library should allow complete

  separation of low-level protocol calls and business logic; by

  comparison, the threaded approach combined with HTTPs protocol

  required (for the first iteration) absolutely no changes from the 1.2

  code, and later changes for optimizing the inter-node RPC calls

  required just syntactic changes (e.g.  ``rpc.call_...`` to

  ``self.rpc.call_...``)

Another issue is with the Twisted API stability - during the Ganeti

1.x lifetime, we had to to implement many times workarounds to changes

in the Twisted version, so that for example 1.2 is able to use both

Twisted 2.x and 8.x.

In the end, since we already had an HTTP server library for the RAPI,

we just reused that for inter-node communication.

Granular locking

~~~~~~~~~~~~~~~~

We want to make sure that multiple operations can run in parallel on a

Ganeti Cluster. In order for this to happen we need to make sure

concurrently run operations don't step on each other toes and break the

cluster.

This design addresses how we are going to deal with locking so that:

- we preserve data coherency

- we prevent deadlocks

- we prevent job starvation

Reaching the maximum possible parallelism is a Non-Goal. We have

identified a set of operations that are currently bottlenecks and need

to be parallelised and have worked on those. In the future it will be

possible to address other needs, thus making the cluster more and more

parallel one step at a time.

This section only talks about parallelising Ganeti level operations, aka

Logical Units, and the locking needed for that. Any other

synchronization lock needed internally by the code is outside its scope.

Library details

+++++++++++++++

The proposed library has these features:

- internally managing all the locks, making the implementation

  transparent from their usage

- automatically grabbing multiple locks in the right order (avoid

  deadlock)

- ability to transparently handle conversion to more granularity

- support asynchronous operation (future goal)

Locking will be valid only on the master node and will not be a

distributed operation. Therefore, in case of master failure, the

operations currently running will be aborted and the locks will be

lost; it remains to the administrator to cleanup (if needed) the

operation result (e.g. make sure an instance is either installed

correctly or removed).

A corollary of this is that a master-failover operation with both

masters alive needs to happen while no operations are running, and

therefore no locks are held.

All the locks will be represented by objects (like

``lockings.SharedLock``), and the individual locks for each object

will be created at initialisation time, from the config file.

The API will have a way to grab one or more than one locks at the same

time.  Any attempt to grab a lock while already holding one in the wrong

order will be checked for, and fail.

The Locks

+++++++++

At the first stage we have decided to provide the following locks:

- One "config file" lock

- One lock per node in the cluster

- One lock per instance in the cluster

All the instance locks will need to be taken before the node locks, and

the node locks before the config lock. Locks will need to be acquired at

the same time for multiple instances and nodes, and internal ordering

will be dealt within the locking library, which, for simplicity, will

just use alphabetical order.

Each lock has the following three possible statuses:

- unlocked (anyone can grab the lock)

- shared (anyone can grab/have the lock but only in shared mode)

- exclusive (no one else can grab/have the lock)

Handling conversion to more granularity

+++++++++++++++++++++++++++++++++++++++

In order to convert to a more granular approach transparently each time

we split a lock into more we'll create a "metalock", which will depend

on those sub-locks and live for the time necessary for all the code to

convert (or forever, in some conditions). When a metalock exists all

converted code must acquire it in shared mode, so it can run

concurrently, but still be exclusive with old code, which acquires it

exclusively.

In the beginning the only such lock will be what replaces the current

"command" lock, and will acquire all the locks in the system, before

proceeding. This lock will be called the "Big Ganeti Lock" because

holding that one will avoid any other concurrent Ganeti operations.

We might also want to devise more metalocks (eg. all nodes, all

nodes+config) in order to make it easier for some parts of the code to

acquire what it needs without specifying it explicitly.

In the future things like the node locks could become metalocks, should

we decide to split them into an even more fine grained approach, but

this will probably be only after the first 2.0 version has been

released.

Adding/Removing locks

+++++++++++++++++++++

When a new instance or a new node is created an associated lock must be

added to the list. The relevant code will need to inform the locking

library of such a change.

This needs to be compatible with every other lock in the system,

especially metalocks that guarantee to grab sets of resources without

specifying them explicitly. The implementation of this will be handled

in the locking library itself.

When instances or nodes disappear from the cluster the relevant locks

must be removed. This is easier than adding new elements, as the code

which removes them must own them exclusively already, and thus deals

with metalocks exactly as normal code acquiring those locks. Any

operation queuing on a removed lock will fail after its removal.

Asynchronous operations

+++++++++++++++++++++++

For the first version the locking library will only export synchronous

operations, which will block till the needed lock are held, and only

fail if the request is impossible or somehow erroneous.

In the future we may want to implement different types of asynchronous

operations such as:

- try to acquire this lock set and fail if not possible

- try to acquire one of these lock sets and return the first one you

  were able to get (or after a timeout) (select/poll like)

These operations can be used to prioritize operations based on available

locks, rather than making them just blindly queue for acquiring them.

The inherent risk, though, is that any code using the first operation,

or setting a timeout for the second one, is susceptible to starvation

and thus may never be able to get the required locks and complete

certain tasks. Considering this providing/using these operations should

not be among our first priorities.

Locking granularity

+++++++++++++++++++

For the first version of this code we'll convert each Logical Unit to

acquire/release the locks it needs, so locking will be at the Logical

Unit level.  In the future we may want to split logical units in

independent "tasklets" with their own locking requirements. A different

design doc (or mini design doc) will cover the move from Logical Units

to tasklets.

Code examples

+++++++++++++

In general when acquiring locks we should use a code path equivalent

to::

  lock.acquire()

  try:

    ...

    # other code

  finally:

    lock.release()

This makes sure we release all locks, and avoid possible deadlocks. Of

course extra care must be used not to leave, if possible locked

structures in an unusable state. Note that with Python 2.5 a simpler

syntax will be possible, but we want to keep compatibility with Python

2.4 so the new constructs should not be used.

In order to avoid this extra indentation and code changes everywhere in

the Logical Units code, we decided to allow LUs to declare locks, and

then execute their code with their locks acquired. In the new world LUs

are called like this::

  # user passed names are expanded to the internal lock/resource name,

  # then known needed locks are declared

  lu.ExpandNames()

  ... some locking/adding of locks may happen ...

  # late declaration of locks for one level: this is useful because sometimes

  # we can't know which resource we need before locking the previous level

  lu.DeclareLocks() # for each level (cluster, instance, node)

  ... more locking/adding of locks can happen ...

  # these functions are called with the proper locks held

  lu.CheckPrereq()

  lu.Exec()

  ... locks declared for removal are removed, all acquired locks released ...

The Processor and the LogicalUnit class will contain exact documentation

on how locks are supposed to be declared.

Caveats

+++++++

This library will provide an easy upgrade path to bring all the code to

granular locking without breaking everything, and it will also guarantee

against a lot of common errors. Code switching from the old "lock

everything" lock to the new system, though, needs to be carefully

scrutinised to be sure it is really acquiring all the necessary locks,

and none has been overlooked or forgotten.

The code can contain other locks outside of this library, to synchronise

other threaded code (eg for the job queue) but in general these should

be leaf locks or carefully structured non-leaf ones, to avoid deadlock

race conditions.

.. _jqueue-original-design:

Job Queue

~~~~~~~~~

Granular locking is not enough to speed up operations, we also need a

queue to store these and to be able to process as many as possible in

parallel.

A Ganeti job will consist of multiple ``OpCodes`` which are the basic

element of operation in Ganeti 1.2 (and will remain as such). Most

command-level commands are equivalent to one OpCode, or in some cases

to a sequence of opcodes, all of the same type (e.g. evacuating a node

will generate N opcodes of type replace disks).

Job execution—“Life of a Ganeti job”

++++++++++++++++++++++++++++++++++++

#. Job gets submitted by the client. A new job identifier is generated

   and assigned to the job. The job is then automatically replicated

   [#replic]_ to all nodes in the cluster. The identifier is returned to

   the client.

#. A pool of worker threads waits for new jobs. If all are busy, the job

   has to wait and the first worker finishing its work will grab it.

   Otherwise any of the waiting threads will pick up the new job.

#. Client waits for job status updates by calling a waiting RPC

   function. Log message may be shown to the user. Until the job is

   started, it can also be canceled.

#. As soon as the job is finished, its final result and status can be

   retrieved from the server.

#. If the client archives the job, it gets moved to a history directory.

   There will be a method to archive all jobs older than a a given age.

.. [#replic] We need replication in order to maintain the consistency

   across all nodes in the system; the master node only differs in the

   fact that now it is running the master daemon, but it if fails and we

   do a master failover, the jobs are still visible on the new master

   (though marked as failed).

Failures to replicate a job to other nodes will be only flagged as

errors in the master daemon log if more than half of the nodes failed,

otherwise we ignore the failure, and rely on the fact that the next

update (for still running jobs) will retry the update. For finished

jobs, it is less of a problem.

Future improvements will look into checking the consistency of the job

list and jobs themselves at master daemon startup.

Job storage

+++++++++++

Jobs are stored in the filesystem as individual files, serialized

using JSON (standard serialization mechanism in Ganeti).

The choice of storing each job in its own file was made because:

- a file can be atomically replaced

- a file can easily be replicated to other nodes

- checking consistency across nodes can be implemented very easily,

  since all job files should be (at a given moment in time) identical

The other possible choices that were discussed and discounted were:

- single big file with all job data: not feasible due to difficult

  updates

- in-process databases: hard to replicate the entire database to the

  other nodes, and replicating individual operations does not mean wee

  keep consistency

Queue structure

+++++++++++++++

All file operations have to be done atomically by writing to a temporary

file and subsequent renaming. Except for log messages, every change in a

job is stored and replicated to other nodes.

::

  /var/lib/ganeti/queue/

    job-1 (JSON encoded job description and status)

    […]

    job-37

    job-38

    job-39

    lock (Queue managing process opens this file in exclusive mode)

    serial (Last job ID used)

    version (Queue format version)

Locking

+++++++

Locking in the job queue is a complicated topic. It is called from more

than one thread and must be thread-safe. For simplicity, a single lock

is used for the whole job queue.

A more detailed description can be found in doc/locking.rst.

Internal RPC

++++++++++++

RPC calls available between Ganeti master and node daemons:

jobqueue_update(file_name, content)

  Writes a file in the job queue directory.

jobqueue_purge()

  Cleans the job queue directory completely, including archived job.

jobqueue_rename(old, new)

  Renames a file in the job queue directory.

Client RPC

++++++++++

RPC between Ganeti clients and the Ganeti master daemon supports the

following operations:

SubmitJob(ops)

  Submits a list of opcodes and returns the job identifier. The

  identifier is guaranteed to be unique during the lifetime of a

  cluster.

WaitForJobChange(job_id, fields, […], timeout)

  This function waits until a job changes or a timeout expires. The

  condition for when a job changed is defined by the fields passed and

  the last log message received.

QueryJobs(job_ids, fields)

  Returns field values for the job identifiers passed.

CancelJob(job_id)

  Cancels the job specified by identifier. This operation may fail if

  the job is already running, canceled or finished.

ArchiveJob(job_id)

  Moves a job into the …/archive/ directory. This operation will fail if

  the job has not been canceled or finished.

Job and opcode status

+++++++++++++++++++++

Each job and each opcode has, at any time, one of the following states:

Queued

  The job/opcode was submitted, but did not yet start.

Waiting

  The job/opcode is waiting for a lock to proceed.

Running

  The job/opcode is running.

Canceled

  The job/opcode was canceled before it started.

Success

  The job/opcode ran and finished successfully.

Error

  The job/opcode was aborted with an error.

If the master is aborted while a job is running, the job will be set to

the Error status once the master started again.

History

+++++++

Archived jobs are kept in a separate directory,

``/var/lib/ganeti/queue/archive/``.  This is done in order to speed up

the queue handling: by default, the jobs in the archive are not

touched by any functions. Only the current (unarchived) jobs are

parsed, loaded, and verified (if implemented) by the master daemon.

Ganeti updates

++++++++++++++

The queue has to be completely empty for Ganeti updates with changes

in the job queue structure. In order to allow this, there will be a

way to prevent new jobs entering the queue.

Object parameters

~~~~~~~~~~~~~~~~~

Across all cluster configuration data, we have multiple classes of

parameters:

A. cluster-wide parameters (e.g. name of the cluster, the master);

   these are the ones that we have today, and are unchanged from the

   current model

#. node parameters

#. instance specific parameters, e.g. the name of disks (LV), that

   cannot be shared with other instances

#. instance parameters, that are or can be the same for many

   instances, but are not hypervisor related; e.g. the number of VCPUs,

   or the size of memory

#. instance parameters that are hypervisor specific (e.g. kernel_path

   or PAE mode)

The following definitions for instance parameters will be used below:

:hypervisor parameter:

  a hypervisor parameter (or hypervisor specific parameter) is defined

  as a parameter that is interpreted by the hypervisor support code in

  Ganeti and usually is specific to a particular hypervisor (like the

  kernel path for :term:`PVM` which makes no sense for :term:`HVM`).

:backend parameter:

  a backend parameter is defined as an instance parameter that can be

  shared among a list of instances, and is either generic enough not

  to be tied to a given hypervisor or cannot influence at all the

  hypervisor behaviour.

  For example: memory, vcpus, auto_balance

  All these parameters will be encoded into constants.py with the prefix

  "BE\_" and the whole list of parameters will exist in the set

  "BES_PARAMETERS"

:proper parameter:

  a parameter whose value is unique to the instance (e.g. the name of a

  LV, or the MAC of a NIC)

As a general rule, for all kind of parameters, “None” (or in

JSON-speak, “nil”) will no longer be a valid value for a parameter. As

such, only non-default parameters will be saved as part of objects in

the serialization step, reducing the size of the serialized format.

Cluster parameters

++++++++++++++++++

Cluster parameters remain as today, attributes at the top level of the

Cluster object. In addition, two new attributes at this level will

hold defaults for the instances:

- hvparams, a dictionary indexed by hypervisor type, holding default

  values for hypervisor parameters that are not defined/overridden by

  the instances of this hypervisor type

- beparams, a dictionary holding (for 2.0) a single element 'default',

  which holds the default value for backend parameters

Node parameters

+++++++++++++++

Node-related parameters are very few, and we will continue using the

same model for these as previously (attributes on the Node object).

There are three new node flags, described in a separate section "node

flags" below.

Instance parameters

+++++++++++++++++++

As described before, the instance parameters are split in three:

instance proper parameters, unique to each instance, instance

hypervisor parameters and instance backend parameters.

The “hvparams” and “beparams” are kept in two dictionaries at instance

level. Only non-default parameters are stored (but once customized, a

parameter will be kept, even with the same value as the default one,

until reset).

The names for hypervisor parameters in the instance.hvparams subtree

should be choosen as generic as possible, especially if specific

parameters could conceivably be useful for more than one hypervisor,

e.g. ``instance.hvparams.vnc_console_port`` instead of using both

``instance.hvparams.hvm_vnc_console_port`` and

``instance.hvparams.kvm_vnc_console_port``.

There are some special cases related to disks and NICs (for example):

a disk has both Ganeti-related parameters (e.g. the name of the LV)

and hypervisor-related parameters (how the disk is presented to/named

in the instance). The former parameters remain as proper-instance

parameters, while the latter value are migrated to the hvparams

structure. In 2.0, we will have only globally-per-instance such

hypervisor parameters, and not per-disk ones (e.g. all NICs will be

exported as of the same type).

Starting from the 1.2 list of instance parameters, here is how they

will be mapped to the three classes of parameters:

- name (P)

- primary_node (P)

- os (P)

- hypervisor (P)

- status (P)

- memory (BE)

- vcpus (BE)

- nics (P)

- disks (P)

- disk_template (P)

- network_port (P)

- kernel_path (HV)

- initrd_path (HV)

- hvm_boot_order (HV)

- hvm_acpi (HV)

- hvm_pae (HV)

- hvm_cdrom_image_path (HV)

- hvm_nic_type (HV)

- hvm_disk_type (HV)

- vnc_bind_address (HV)

- serial_no (P)

Parameter validation

++++++++++++++++++++

To support the new cluster parameter design, additional features will

be required from the hypervisor support implementations in Ganeti.

The hypervisor support  implementation API will be extended with the

following features:

:PARAMETERS: class-level attribute holding the list of valid parameters

  for this hypervisor

:CheckParamSyntax(hvparams): checks that the given parameters are

  valid (as in the names are valid) for this hypervisor; usually just

  comparing ``hvparams.keys()`` and ``cls.PARAMETERS``; this is a class

  method that can be called from within master code (i.e. cmdlib) and

  should be safe to do so

:ValidateParameters(hvparams): verifies the values of the provided

  parameters against this hypervisor; this is a method that will be

  called on the target node, from backend.py code, and as such can

  make node-specific checks (e.g. kernel_path checking)

Default value application

+++++++++++++++++++++++++

The application of defaults to an instance is done in the Cluster

object, via two new methods as follows:

- ``Cluster.FillHV(instance)``, returns 'filled' hvparams dict, based on

  instance's hvparams and cluster's ``hvparams[instance.hypervisor]``

- ``Cluster.FillBE(instance, be_type="default")``, which returns the

  beparams dict, based on the instance and cluster beparams

The FillHV/BE transformations will be used, for example, in the

RpcRunner when sending an instance for activation/stop, and the sent

instance hvparams/beparams will have the final value (noded code doesn't

know about defaults).

LU code will need to self-call the transformation, if needed.

Opcode changes

++++++++++++++

The parameter changes will have impact on the OpCodes, especially on

the following ones:

- ``OpInstanceCreate``, where the new hv and be parameters will be sent

  as dictionaries; note that all hv and be parameters are now optional,

  as the values can be instead taken from the cluster

- ``OpInstanceQuery``, where we have to be able to query these new

  parameters; the syntax for names will be ``hvparam/$NAME`` and

  ``beparam/$NAME`` for querying an individual parameter out of one

  dictionary, and ``hvparams``, respectively ``beparams``, for the whole

  dictionaries

- ``OpModifyInstance``, where the the modified parameters are sent as

  dictionaries

Additionally, we will need new OpCodes to modify the cluster-level

defaults for the be/hv sets of parameters.

Caveats

+++++++

One problem that might appear is that our classification is not

complete or not good enough, and we'll need to change this model. As

the last resort, we will need to rollback and keep 1.2 style.

Another problem is that classification of one parameter is unclear

(e.g. ``network_port``, is this BE or HV?); in this case we'll take

the risk of having to move parameters later between classes.

Security

++++++++

The only security issue that we foresee is if some new parameters will

have sensitive value. If so, we will need to have a way to export the

config data while purging the sensitive value.

E.g. for the drbd shared secrets, we could export these with the

values replaced by an empty string.

Node flags

~~~~~~~~~~

Ganeti 2.0 adds three node flags that change the way nodes are handled

within Ganeti and the related infrastructure (iallocator interaction,

RAPI data export).

*master candidate* flag

+++++++++++++++++++++++

Ganeti 2.0 allows more scalability in operation by introducing

parallelization. However, a new bottleneck is reached that is the

synchronization and replication of cluster configuration to all nodes

in the cluster.

This breaks scalability as the speed of the replication decreases

roughly with the size of the nodes in the cluster. The goal of the

master candidate flag is to change this O(n) into O(1) with respect to

job and configuration data propagation.

Only nodes having this flag set (let's call this set of nodes the

*candidate pool*) will have jobs and configuration data replicated.

The cluster will have a new parameter (runtime changeable) called

``candidate_pool_size`` which represents the number of candidates the

cluster tries to maintain (preferably automatically).

This will impact the cluster operations as follows:

- jobs and config data will be replicated only to a fixed set of nodes

- master fail-over will only be possible to a node in the candidate pool

- cluster verify needs changing to account for these two roles

- external scripts will no longer have access to the configuration

  file (this is not recommended anyway)

The caveats of this change are:

- if all candidates are lost (completely), cluster configuration is

  lost (but it should be backed up external to the cluster anyway)

- failed nodes which are candidate must be dealt with properly, so

  that we don't lose too many candidates at the same time; this will be

  reported in cluster verify

- the 'all equal' concept of ganeti is no longer true

- the partial distribution of config data means that all nodes will

  have to revert to ssconf files for master info (as in 1.2)

Advantages:

- speed on a 100+ nodes simulated cluster is greatly enhanced, even

  for a simple operation; ``gnt-instance remove`` on a diskless instance

  remove goes from ~9seconds to ~2 seconds

- node failure of non-candidates will be less impacting on the cluster

The default value for the candidate pool size will be set to 10 but

this can be changed at cluster creation and modified any time later.

Testing on simulated big clusters with sequential and parallel jobs

show that this value (10) is a sweet-spot from performance and load

point of view.

*offline* flag

++++++++++++++

In order to support better the situation in which nodes are offline

(e.g. for repair) without altering the cluster configuration, Ganeti

needs to be told and needs to properly handle this state for nodes.

This will result in simpler procedures, and less mistakes, when the

amount of node failures is high on an absolute scale (either due to

high failure rate or simply big clusters).

Nodes having this attribute set will not be contacted for inter-node

RPC calls, will not be master candidates, and will not be able to host

instances as primaries.

Setting this attribute on a node:

- will not be allowed if the node is the master

- will not be allowed if the node has primary instances

- will cause the node to be demoted from the master candidate role (if

  it was), possibly causing another node to be promoted to that role

This attribute will impact the cluster operations as follows:

- querying these nodes for anything will fail instantly in the RPC

  library, with a specific RPC error (RpcResult.offline == True)

- they will be listed in the Other section of cluster verify

The code is changed in the following ways:

- RPC calls were be converted to skip such nodes:

  - RpcRunner-instance-based RPC calls are easy to convert

  - static/classmethod RPC calls are harder to convert, and were left

    alone

- the RPC results were unified so that this new result state (offline)

  can be differentiated

- master voting still queries in repair nodes, as we need to ensure

  consistency in case the (wrong) masters have old data, and nodes have

  come back from repairs

Caveats:

- some operation semantics are less clear (e.g. what to do on instance

  start with offline secondary?); for now, these will just fail as if

  the flag is not set (but faster)

- 2-node cluster with one node offline needs manual startup of the

  master with a special flag to skip voting (as the master can't get a

  quorum there)

One of the advantages of implementing this flag is that it will allow

in the future automation tools to automatically put the node in

repairs and recover from this state, and the code (should/will) handle

this much better than just timing out. So, future possible

improvements (for later versions):

- watcher will detect nodes which fail RPC calls, will attempt to ssh

  to them, if failure will put them offline

- watcher will try to ssh and query the offline nodes, if successful

  will take them off the repair list

Alternatives considered: The RPC call model in 2.0 is, by default,

much nicer - errors are logged in the background, and job/opcode

execution is clearer, so we could simply not introduce this. However,

having this state will make both the codepaths clearer (offline

vs. temporary failure) and the operational model (it's not a node with

errors, but an offline node).

*drained* flag

++++++++++++++

Due to parallel execution of jobs in Ganeti 2.0, we could have the

following situation:

- gnt-node migrate + failover is run

- gnt-node evacuate is run, which schedules a long-running 6-opcode

  job for the node

- partway through, a new job comes in that runs an iallocator script,

  which finds the above node as empty and a very good candidate

- gnt-node evacuate has finished, but now it has to be run again, to

  clean the above instance(s)

In order to prevent this situation, and to be able to get nodes into

proper offline status easily, a new *drained* flag was added to the

nodes.

This flag (which actually means "is being, or was drained, and is

expected to go offline"), will prevent allocations on the node, but

otherwise all other operations (start/stop instance, query, etc.) are

working without any restrictions.

Interaction between flags

+++++++++++++++++++++++++

While these flags are implemented as separate flags, they are

mutually-exclusive and are acting together with the master node role

as a single *node status* value. In other words, a flag is only in one

of these roles at a given time. The lack of any of these flags denote

a regular node.

The current node status is visible in the ``gnt-cluster verify``

output, and the individual flags can be examined via separate flags in

the ``gnt-node list`` output.

These new flags will be exported in both the iallocator input message

and via RAPI, see the respective man pages for the exact names.

Feature changes

---------------

The main feature-level changes will be:

- a number of disk related changes

- removal of fixed two-disk, one-nic per instance limitation

Disk handling changes

~~~~~~~~~~~~~~~~~~~~~

The storage options available in Ganeti 1.x were introduced based on

then-current software (first DRBD 0.7 then later DRBD 8) and the

estimated usage patters. However, experience has later shown that some

assumptions made initially are not true and that more flexibility is

needed.

One main assumption made was that disk failures should be treated as

'rare' events, and that each of them needs to be manually handled in

order to ensure data safety; however, both these assumptions are false:

- disk failures can be a common occurrence, based on usage patterns or

  cluster size

- our disk setup is robust enough (referring to DRBD8 + LVM) that we

  could automate more of the recovery

Note that we still don't have fully-automated disk recovery as a goal,

but our goal is to reduce the manual work needed.

As such, we plan the following main changes:

- DRBD8 is much more flexible and stable than its previous version

  (0.7), such that removing the support for the ``remote_raid1``

  template and focusing only on DRBD8 is easier

- dynamic discovery of DRBD devices is not actually needed in a cluster

  that where the DRBD namespace is controlled by Ganeti; switching to a

  static assignment (done at either instance creation time or change

  secondary time) will change the disk activation time from O(n) to

  O(1), which on big clusters is a significant gain

- remove the hard dependency on LVM (currently all available storage

  types are ultimately backed by LVM volumes) by introducing file-based

  storage

Additionally, a number of smaller enhancements are also planned:

- support variable number of disks

- support read-only disks

Future enhancements in the 2.x series, which do not require base design

changes, might include:

- enhancement of the LVM allocation method in order to try to keep

  all of an instance's virtual disks on the same physical

  disks

- add support for DRBD8 authentication at handshake time in

  order to ensure each device connects to the correct peer

- remove the restrictions on failover only to the secondary

  which creates very strict rules on cluster allocation

DRBD minor allocation

+++++++++++++++++++++

Currently, when trying to identify or activate a new DRBD (or MD)

device, the code scans all in-use devices in order to see if we find

one that looks similar to our parameters and is already in the desired

state or not. Since this needs external commands to be run, it is very

slow when more than a few devices are already present.

Therefore, we will change the discovery model from dynamic to

static. When a new device is logically created (added to the

configuration) a free minor number is computed from the list of

devices that should exist on that node and assigned to that

device.

At device activation, if the minor is already in use, we check if

it has our parameters; if not so, we just destroy the device (if

possible, otherwise we abort) and start it with our own

parameters.

This means that we in effect take ownership of the minor space for

that device type; if there's a user-created DRBD minor, it will be

automatically removed.

The change will have the effect of reducing the number of external

commands run per device from a constant number times the index of the

first free DRBD minor to just a constant number.

Removal of obsolete device types (MD, DRBD7)

++++++++++++++++++++++++++++++++++++++++++++

We need to remove these device types because of two issues. First,

DRBD7 has bad failure modes in case of dual failures (both network and

disk - it cannot propagate the error up the device stack and instead

just panics. Second, due to the asymmetry between primary and

secondary in MD+DRBD mode, we cannot do live failover (not even if we

had MD+DRBD8).

File-based storage support

++++++++++++++++++++++++++

Using files instead of logical volumes for instance storage would

allow us to get rid of the hard requirement for volume groups for

testing clusters and it would also allow usage of SAN storage to do

live failover taking advantage of this storage solution.

Better LVM allocation

+++++++++++++++++++++

Currently, the LV to PV allocation mechanism is a very simple one: at

each new request for a logical volume, tell LVM to allocate the volume

in order based on the amount of free space. This is good for

simplicity and for keeping the usage equally spread over the available

physical disks, however it introduces a problem that an instance could

end up with its (currently) two drives on two physical disks, or

(worse) that the data and metadata for a DRBD device end up on

different drives.

This is bad because it causes unneeded ``replace-disks`` operations in

case of a physical failure.

The solution is to batch allocations for an instance and make the LVM

handling code try to allocate as close as possible all the storage of

one instance. We will still allow the logical volumes to spill over to

additional disks as needed.

Note that this clustered allocation can only be attempted at initial

instance creation, or at change secondary node time. At add disk time,

or at replacing individual disks, it's not easy enough to compute the

current disk map so we'll not attempt the clustering.

DRBD8 peer authentication at handshake

++++++++++++++++++++++++++++++++++++++

DRBD8 has a new feature that allow authentication of the peer at

connect time. We can use this to prevent connecting to the wrong peer

more that securing the connection. Even though we never had issues

with wrong connections, it would be good to implement this.

LVM self-repair (optional)

++++++++++++++++++++++++++

The complete failure of a physical disk is very tedious to

troubleshoot, mainly because of the many failure modes and the many

steps needed. We can safely automate some of the steps, more

specifically the ``vgreduce --removemissing`` using the following

method:

#. check if all nodes have consistent volume groups

#. if yes, and previous status was yes, do nothing

#. if yes, and previous status was no, save status and restart

#. if no, and previous status was no, do nothing

#. if no, and previous status was yes:

    #. if more than one node is inconsistent, do nothing

    #. if only one node is inconsistent:

        #. run ``vgreduce --removemissing``

        #. log this occurrence in the Ganeti log in a form that

           can be used for monitoring

        #. [FUTURE] run ``replace-disks`` for all

           instances affected

Failover to any node

++++++++++++++++++++

With a modified disk activation sequence, we can implement the

*failover to any* functionality, removing many of the layout

restrictions of a cluster:

- the need to reserve memory on the current secondary: this gets reduced

  to a must to reserve memory anywhere on the cluster

- the need to first failover and then replace secondary for an

  instance: with failover-to-any, we can directly failover to

  another node, which also does the replace disks at the same

  step

In the following, we denote the current primary by P1, the current

secondary by S1, and the new primary and secondaries by P2 and S2. P2

is fixed to the node the user chooses, but the choice of S2 can be

made between P1 and S1. This choice can be constrained, depending on

which of P1 and S1 has failed.

- if P1 has failed, then S1 must become S2, and live migration is not

  possible

- if S1 has failed, then P1 must become S2, and live migration could be

  possible (in theory, but this is not a design goal for 2.0)

The algorithm for performing the failover is straightforward:

- verify that S2 (the node the user has chosen to keep as secondary) has

  valid data (is consistent)

- tear down the current DRBD association and setup a DRBD pairing

  between P2 (P2 is indicated by the user) and S2; since P2 has no data,

  it will start re-syncing from S2

- as soon as P2 is in state SyncTarget (i.e. after the resync has

  started but before it has finished), we can promote it to primary role

  (r/w) and start the instance on P2

- as soon as the P2?S2 sync has finished, we can remove

  the old data on the old node that has not been chosen for

  S2

Caveats: during the P2?S2 sync, a (non-transient) network error

will cause I/O errors on the instance, so (if a longer instance

downtime is acceptable) we can postpone the restart of the instance

until the resync is done. However, disk I/O errors on S2 will cause

data loss, since we don't have a good copy of the data anymore, so in

this case waiting for the sync to complete is not an option. As such,

it is recommended that this feature is used only in conjunction with

proper disk monitoring.

Live migration note: While failover-to-any is possible for all choices

of S2, migration-to-any is possible only if we keep P1 as S2.

Caveats

+++++++

The dynamic device model, while more complex, has an advantage: it

will not reuse by mistake the DRBD device of another instance, since

it always looks for either our own or a free one.

The static one, in contrast, will assume that given a minor number N,

it's ours and we can take over. This needs careful implementation such

that if the minor is in use, either we are able to cleanly shut it

down, or we abort the startup. Otherwise, it could be that we start

syncing between two instance's disks, causing data loss.

Variable number of disk/NICs per instance

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Variable number of disks

++++++++++++++++++++++++

In order to support high-security scenarios (for example read-only sda

and read-write sdb), we need to make a fully flexibly disk

definition. This has less impact that it might look at first sight:

only the instance creation has hard coded number of disks, not the disk

handling code. The block device handling and most of the instance

handling code is already working with "the instance's disks" as

opposed to "the two disks of the instance", but some pieces are not

(e.g. import/export) and the code needs a review to ensure safety.

The objective is to be able to specify the number of disks at

instance creation, and to be able to toggle from read-only to

read-write a disk afterward.

Variable number of NICs

+++++++++++++++++++++++

Similar to the disk change, we need to allow multiple network

interfaces per instance. This will affect the internal code (some

function will have to stop assuming that ``instance.nics`` is a list

of length one), the OS API which currently can export/import only one

instance, and the command line interface.

Interface changes

-----------------

There are two areas of interface changes: API-level changes (the OS

interface and the RAPI interface) and the command line interface

changes.

OS interface

~~~~~~~~~~~~

The current Ganeti OS interface, version 5, is tailored for Ganeti 1.2.

The interface is composed by a series of scripts which get called with

certain parameters to perform OS-dependent operations on the cluster.

The current scripts are:

create

  called when a new instance is added to the cluster

export

  called to export an instance disk to a stream

import

  called to import from a stream to a new instance

rename

  called to perform the os-specific operations necessary for renaming an

  instance

Currently these scripts suffer from the limitations of Ganeti 1.2: for

example they accept exactly one block and one swap devices to operate

on, rather than any amount of generic block devices, they blindly assume

that an instance will have just one network interface to operate, they

can not be configured to optimise the instance for a particular

hypervisor.

Since in Ganeti 2.0 we want to support multiple hypervisors, and a

non-fixed number of network and disks the OS interface need to change to

transmit the appropriate amount of information about an instance to its

managing operating system, when operating on it. Moreover since some old

assumptions usually used in OS scripts are no longer valid we need to

re-establish a common knowledge on what can be assumed and what cannot

be regarding Ganeti environment.

When designing the new OS API our priorities are:

- ease of use

- future extensibility

- ease of porting from the old API

- modularity

As such we want to limit the number of scripts that must be written to

support an OS, and make it easy to share code between them by uniforming

their input.  We also will leave the current script structure unchanged,

as far as we can, and make a few of the scripts (import, export and

rename) optional. Most information will be passed to the script through

environment variables, for ease of access and at the same time ease of

using only the information a script needs.

The Scripts

+++++++++++

As in Ganeti 1.2, every OS which wants to be installed in Ganeti needs

to support the following functionality, through scripts:

create:

  used to create a new instance running that OS. This script should

  prepare the block devices, and install them so that the new OS can

  boot under the specified hypervisor.

export (optional):