open-consul/website/source/docs/internals/sessions.html.markdown

116 lines
6.1 KiB
Markdown

---
layout: "docs"
page_title: "Sessions"
sidebar_current: "docs-internals-sessions"
description: |-
Consul provides a session mechanism which can be used to build distributed locks. Sessions act as a binding layer between nodes, health checks, and key/value data. They are designed to provide granular locking, and are heavily inspired by The Chubby Lock Service for Loosely-Coupled Distributed Systems.
---
# Sessions
Consul provides a session mechanism which can be used to build distributed locks.
Sessions act as a binding layer between nodes, health checks, and key/value data.
They are designed to provide granular locking, and are heavily inspired
by [The Chubby Lock Service for Loosely-Coupled Distributed Systems](http://research.google.com/archive/chubby.html).
~> **Advanced Topic!** This page covers technical details of
the internals of Consul. You don't need to know these details to effectively
operate and use Consul. These details are documented here for those who wish
to learn about them without having to go spelunking through the source code.
## Session Design
A session in Consul represents a contract that has very specific semantics.
When a session is constructed a node name, a list of health checks, and a
`lock-delay` are provided. The newly constructed session is provided with
a named ID which can be used to refer to it. This ID can be used with the KV
store to acquire locks, which are advisory mechanisms for mutual exclusion.
Below is a diagram showing the relationship between these components:
<div class="center">
![Consul Sessions](consul-sessions.png)
</div>
The contract that Consul provides is that under any of the following
situations the session will be *invalidated*:
* Node is deregistered
* Any of the health checks are deregistered
* Any of the health checks go to the critical state
* Session is explicitly destroyed
When a session is invalidated, any of the locks held in association
with the session are released, and the `ModifyIndex` of the key is
incremented. The session is also destroyed during an invalidation
and can no longer be used to acquire further locks.
While this is a simple design, it enables a multitude of usage
patterns. By default, the [gossip based failure detector](/docs/internals/gossip.html)
is used as the associated health check. This failure detector allows
Consul to detect when a node that is holding a lock has failed, and
to automatically release the lock. This ability provides **liveness** to
Consul locks, meaning under failure the system can continue to make
progress. However, because there is no perfect failure detector, it's possible
to have a false positive (failure detected) which causes the lock to
be released even though the lock owner is still alive. This means
we are sacrificing some **safety**.
Conversely, it is possible to create a session with no associated
health checks. This removes the possibility of a false positive,
and trades liveness for safety. You can be absolutely certain Consul
will not release the lock even if the existing owner has failed.
Since Consul APIs allow a session to be force destroyed, this allows
systems to be built that require an operator to intervene in the
case of a failure, but preclude the possibility of a split-brain.
The final nuance is that sessions may provide a `lock-delay`. This
is a time duration, between 0 and 60 seconds. When a session invalidation
takes place, Consul prevents any of the previously held locks from
being re-acquired for the `lock-delay` interval; this is a safeguard
inspired by Google's Chubby. The purpose of this delay is to allow
the potentially still live leader to detect the invalidation and stop
processing requests that may lead to inconsistent state. While not a
bulletproof method, it does avoid the need to introduce sleep states
into application logic, and can help mitigate many issues. While the
default is to use a 15 second delay, clients are able to disable this
mechanism by providing a zero delay value.
## KV Integration
Integration between the Key/Value store and sessions are the primary
place where sessions are used. A session must be created prior to use,
and is then referred to by its ID.
The Key/Value API is extended to support an `acquire` and `release` operation.
The `acquire` operation acts like a Check-And-Set operation, except it
can only succeed if there is no existing lock holder. On success, there
is a normal key update, but there is also an increment to the `LockIndex`,
and the `Session` value is updated to reflect the session holding the lock.
Once held, the lock can be released using a corresponding `release` operation,
providing the same session. Again, this acts like a Check-And-Set operations,
since the request will fail if given an invalid session. A critical note is
that the lock can be released without being the creator of the session.
This is by design, as it allows operators to intervene and force terminate
a session if necessary. As mentioned above, a session invalidation will also
cause all held locks to be released. When a lock is released, the `LockIndex`
does not change, however the `Session` is cleared and the `ModifyIndex` increments.
These semantics (heavily borrowed from Chubby), allow the tuple of (Key, LockIndex, Session)
to act as a unique "sequencer". This `sequencer` can be passed around and used
to verify if the request belongs to the current lock holder. Because the `LockIndex`
is incremented on each `acquire`, even if the same session re-acquires a lock,
the `sequencer` will be able to detect a stale request. Similarly, if a session is
invalided, the Session corresponding to the given `LockIndex` will be blank.
To be clear, this locking system is purely *advisory*. There is no enforcement
that clients must acquire a lock to perform any operation. Any client can
read, write, and delete a key without owning the corresponding lock. It is not
the goal of Consul to protect against misbehaving clients.
## Leader Election
The primitives provided by sessions and the locking mechanisms of the KV
store can be used to build client-side leader election algorithms.
These are covered in more detail in the [Leader Election guide](/docs/guides/leader-election.html).