tursom/memgraph
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Update HA feature spec

Browse Source 
Reviewers: buda, teon.banek, mferencevic

Reviewed By: buda

Subscribers: dsantl, dlozic, tlastre

Differential Revision: https://phabricator.memgraph.io/D2626
This commit is contained in:
Ivan Paljak 2020-01-20 17:53:02 +01:00
parent 3a88668ace
commit 3121f7d89d
1 changed files with 232 additions and 171 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

403
docs/feature_spec/high_availability.md
View File

@ -1,214 +1,275 @@
# High Availability (abbr. HA)
## Introduction
## High Level Context
High availability is a characteristic of a system which aims to ensure a
certain level of operational performance for a higher-than-normal period.
Although there are multiple ways to design highly available systems, Memgraph
strives to achieve HA by elimination of single points of failure. In essence,
this implies adding redundancy to the system so that a failure of a component
does not imply the failure of the entire system.
does not imply the failure of the entire system. To ensure this, HA Memgraph
implements the [Raft consensus algorithm](https://raft.github.io/).
## Theoretical Background
Correct implementation of the algorithm guarantees that the cluster will be
fully functional (available) as long as any strong majority of the servers are
operational and can communicate with each other and with clients. For example,
clusters of three or four machines can tolerate the failure of a single server,
clusters of five and six machines can tolerate the failure of any two servers,
and so on. Therefore, we strongly recommend a setup of an odd-sized cluster.
The following chapter serves as an introduction into some theoretical aspects
of Memgraph's high availability implementation. If the reader is solely
interested in design decisions around HA implementation, they can skip this
chapter.
### Performance Implications
An important implication of any HA implementation stems from Eric Brewer's
[CAP Theorem](https://fenix.tecnico.ulisboa.pt/downloadFile/1126518382178117/10.e-CAP-3.pdf)
which states that it is impossible for a distributed system to simultaneously
achieve:
Internally, Raft achieves high availability by keeping a consistent replicated
log on each server within the cluster. Therefore, we must successfully replicate
a transaction on the majority of servers within the cluster before we actually
commit it and report the result back to the client. This operation represents
a significant performance hit when compared with single node version of
Memgraph.
* Consistency (C - every read receives the most recent write or an error)
* Availability (A - every request receives a response that is not an error)
* Partition tolerance (P - The system continues to operate despite an
arbitrary number of messages being dropped by the
network between nodes)
Luckily, the algorithm can be tweaked in a way which allows read-only
transactions to perform significantly better than those which modify the
database state. That being said, the performance of read-only operations
is still not going to be on par with single node Memgraph.
In the context of HA, Memgraph should strive to achieve CA.
This section will be updated with exact numbers once we integrate HA with
new storage.
### Consensus
With the old storage, write throughput was almost five times lower than read
throughput (~30000 reads per second vs ~6000 writes per second).
Implications of the CAP theorem naturally lead us towards introducing a
cluster of machines which will have identical internal states. When a designated
machine for handling client requests fails, it can simply be replaced with
another.
## User Facing Setup
Well... turns out this is not as easy as it sounds :(
### How to Setup HA Memgraph Cluster?
Keeping around a cluster of machines with consistent internal state is an
inherently difficult problem. More precisely, this problem is as hard as
getting a cluster of machines to agree on a single value, which is a highly
researched area in distributed systems. Our research of state of the art
consensus algorithms lead us to Diego Ongaro's
[Raft algorithm](https://raft.github.io/raft.pdf).
First, the user needs to install `memgraph_ha` package on each machine
in their cluster. HA Memgraph should be available as a Debian package,
so its installation on each machine should be as simple as:
#### Raft
```plaintext
dpkg -i /path/to/memgraph_ha_<version>.deb
```
As you might have guessed, analyzing each subtle detail of Raft goes way
beyond the scope of this document. In the remainder of the chapter we will
outline only the most important ideas and implications, leaving all further
analysis to the reader. Detailed explanation can be found either in Diego's
[dissertation](https://ramcloud.stanford.edu/~ongaro/thesis.pdf) \[1\] or the
Raft [paper](https://raft.github.io/raft.pdf) \[2\].
After successful installation of the `memgraph_ha` package, the user should
finish its configuration before attempting to start the cluster.
In essence, Raft allows us to implement the previously mentioned idea of
managing a cluster of machines with identical internal states. In other
words, the Raft protocol allows us to manage a cluster of replicated
state machines which is fully functional as long as the *majority* of
the machines in the cluster operate correctly.
There are two main things that need to be configured on every node in order for
the cluster to be able to run:
Another important fact is that those state machines must be *deterministic*.
In other words, the same command on two different machines with the same
internal state must yield the same result. This is important because Memgraph,
as a black box, is not entirely deterministic. Non-determinism can easily be
introduced by the user (e.g. by using the `rand` function) or by algorithms
behind query execution (e.g. introducing fuzzy logic in the planner could yield
a different order of results). Luckily, once we enter the storage level,
everything should be fully deterministic.
1. The user has to edit the main configuration file and specify the unique node
ID to each server in the cluster
2. The user has to create a file that describes all IP addresses of all servers
that will be used in the cluster
To summarize, Raft is a protocol which achieves consensus in a cluster of
deterministic state machines via log replication. The cluster is fully
functional if the majority of the machines work correctly. The reader
is strongly encouraged to gain a deeper understanding (at least read through
the paper) of Raft before reading the rest of this document.
The `memgraph_ha` binary loads all main configuration parameters from
`/etc/memgraph/memgraph_ha.conf`. On each node of the cluster, the user should
uncomment the `--server-id=0` parameter and change its value to the `server_id`
of that node.
The last step before starting the server is to create a `coordination`
configuration file. That file is already present as an example in
`/etc/memgraph/coordination.json.example` and you have to copy it to
`/etc/memgraph/coordination.json` and edit it according to your cluster
configuration. The file contains coordination info consisting of a list of
`server_id`, `ip_address` and `rpc_port` lists. The assumed contents of the
`coordination.json` file are:
```plaintext
[
[1, "192.168.0.1", 10000],
[2, "192.168.0.2", 10000],
[3, "192.168.0.3", 10000]
]
```
Here, each line corresponds to coordination of one server. The first entry is
that server's ID, the second is its IP address and the third is the RPC port it
listens to. This port should not be confused with the port used for client
interaction via the Bolt protocol.
The `ip_address` entered for each `server_id` *must* match the exact IP address
that belongs to that server and that will be used to communicate to other nodes
in the cluster. The coordination configuration file *must* be identical on all
nodes in the cluster.
After the user has set the `server_id` on each node in
`/etc/memgraph/memgraph_ha.conf` and provided the same
`/etc/memgraph/coordination.json` file to each node in the cluster, they can
start the Memgraph HA service by issuing the following command on each node in
the cluster:
```plaintext
systemctl start memgraph_ha
```
### How to Configure Raft Parameters?
All Raft configuration parameters can be controlled by modifying
`/etc/memgraph/raft.json`. The assumed contents of the `raft.json` file are:
```plaintext
{
"election_timeout_min": 750,
"election_timeout_max": 1000,
"heartbeat_interval": 100,
"replication_timeout": 20000,
"log_size_snapshot_threshold": 50000
}
```
The meaning behind each entry is demystified in the following table:
Flag | Description
------------------------------|------------
`election_timeout_min` | Lower bound for the randomly sampled reelection timer given in milliseconds
`election_timeout_max` | Upper bound for the randomly sampled reelection timer given in milliseconds
`heartbeat_interval` | Time interval between consecutive heartbeats given in milliseconds
`replication_timeout` | Time interval allowed for data replication given in milliseconds
`log_size_snapshot_threshold` | Allowed number of entries in Raft log before its compaction
### How to Query HA Memgraph via Proxy?
This chapter describes how to query HA Memgraph using our proxy server.
Note that this is not intended to be a long-term solution. Instead, we will
implement a proper Memgraph HA client which is capable of communicating with
the HA cluster. Once our own client is implemented, it will no longer be
possible to query HA Memgraph using other clients (such as neo4j client).
The Bolt protocol that is exposed by each Memgraph HA node is an extended
version of the standard Bolt protocol. In order to be able to communicate with
the highly available cluster of Memgraph HA nodes, the client must have some
logic implemented in itself so that it can communicate correctly with all nodes
in the cluster. To facilitate a faster start with the HA cluster we will build
the Memgraph HA proxy binary that communicates with all nodes in the HA cluster
using the extended Bolt protocol and itself exposes a standard Bolt protocol to
the user. All standard Bolt clients (libraries and custom systems) can
communicate with the Memgraph HA proxy without any code modifications.
The HA proxy should be deployed on each client machine that is used to
communicate with the cluster. It can't be deployed on the Memgraph HA nodes!
When using the Memgraph HA proxy, the communication flow is described in the
following diagram:
```plaintext
Memgraph HA node 1 -----+
|
Memgraph HA node 2 -----+ Memgraph HA proxy <---> any standard Bolt client (C, Java, PHP, Python, etc.)
|
Memgraph HA node 3 -----+
```
To setup the Memgraph HA proxy the user should install the `memgraph_ha_proxy`
package.
After its successful installation, the user should enter all endpoints of the
HA Memgraph cluster servers into the configuration before attempting to start
the HA Memgraph proxy server.
The HA Memgraph proxy server loads all of its configuration from
`/etc/memgraph/memgraph_ha_proxy.conf`. Assuming that the cluster is set up
like in the previous examples, the user should uncomment and enter the following
value into the `--endpoints` parameter:
```plaintext
--endpoints=192.168.0.1:7687,192.168.0.2:7687,192.168.0.3:7687
```
Note that the IP addresses used in the example match the individual cluster
nodes IP addresses, but the ports used are the Bolt server ports exposed by
each node (currently the default value of `7687`).
The user can now start the proxy by using the following command:
```plaintext
systemctl start memgraph_ha_proxy
```
After the proxy has been started, the user can query the HA cluster by
connecting to the HA Memgraph proxy IP address using their favorite Bolt
client.
## Integration with Memgraph
The first thing that should be defined is a single instruction within the
context of Raft (i.e. a single entry in a replicated log). As mentioned
before, these instructions should be completely deterministic when applied
context of Raft (i.e. a single entry in a replicated log).
These instructions should be completely deterministic when applied
to the state machine. We have therefore decided that the appropriate level
of abstraction within Memgraph corresponds to `StateDelta`-s (data structures
of abstraction within Memgraph corresponds to `Delta`s (data structures
which describe a single change to the Memgraph state, used for durability
in WAL). Moreover, a single instruction in a replicated log will consist of a
batch of `StateDelta`s which correspond to a single **committed** transaction.
This decision both improves performance and handles some special cases that
present themselves otherwise by leveraging the knowledge that the transaction
should be committed.
batch of `Delta`s which correspond to a single transaction that's about
to be **committed**.
"What happens with aborted transactions?"
Apart from `Delta`s, there are certain operations within the storage called
`StorageGlobalOperations` which do not conform to usual transactional workflow
(e.g. Creating indices). Since our storage engine implementation guarantees
that at the moment of their execution no other transactions are active, we can
safely replicate them as well. In other words, no additional logic needs to be
implemented because of them.
A great question, they are handled solely by the leader which is the only
machine that communicates with the client. Aborted transactions do not alter
the state of the database and there is no need to replicate it to other machines
in the cluster. If, for instance, the leader dies before returning the result
of some read operation in an aborted transaction, the client will notice that
the leader has crashed. A new leader will be elected in the next term and the
client should retry the transaction.
Therefore, we will introduce a new `RaftDelta` object which can be constructed
both from storage `Delta` and `StorageGlobalOperation`. Instead of appending
these to WAL (as we do in single node), we will start to replicate them across
our cluster. Once we have replicated the corresponding Raft log entry on
majority of the cluster, we are able to safely commit the transaction or execute
a global operation. If for any reason the replication fails (leadership change,
worker failures, etc.) the transaction will be aborted.
"OK, that makes sense! But, wait a minute, this is broken by design! Merely
generating `StateDelta`s on the leader for any transaction will taint its
internal storage before sending the first RPC to some follower. This deviates
from Raft and will crash the universe!"
In the follower mode, we need to be able to apply `RaftDelta`s we got from
the leader when the protocol allows us to do so. In that case, we will use the
same concepts from durability in storage v2, i.e., applying deltas maps
completely to recovery from WAL in storage v2.
Another great observation. It is indeed true that applying `StateDelta`s makes
changes to local storage, but only a single type of `StateDelta` makes that
change durable. That `StateDelta` type is called `TRANSACTION_COMMIT` and we
will change its behaviour when working as a HA instance. More precisely, we
must not allow the transaction engine to modify the commit log saying that
the transaction has been committed. That action should be delayed until those
`StateDelta`s have been applied to the majority of the cluster. At that point
the commit log can be safely modified leaving it up to Raft to ensure the
durability of the transaction.
## Test and Benchmark Strategy
We should also address one subtle detail that arises in this case. Consider
the following scenario:
We have already implemented some integration and stress tests. These are:
* The leader starts working on a transaction which creates a new record in the
database. Suppose that record is stored in the leader's internal storage
but the transaction was not committed (i.e. no such entry in the commit log).
* The leader should start replicating those `StateDelta`s to its followers
but, suddenly, it's cut off from the rest of the cluster.
* Due to timeout, a new election is held and a new leader has been elected.
* Our old leader comes back to life and becomes a follower.
* The new leader receives a transaction which creates that same record, but
this transaction is successfully replicated and committed by the new leader.
1. leader election -- Tests whether leader election works properly.
2. basic test -- Tests basic leader election and log replication.
3. term updates test -- Tests a specific corner case (which used to fail)
regarding term updates.
4. log compaction test -- Tests whether log compaction works properly.
5. large log entries -- Tests whether we can successfully replicate relatively
large log entries.
6. index test -- Tests whether index creation works in HA.
7. normal operation stress test -- Long running concurrent stress test under
normal conditions (no failures).
8. read benchmark -- Measures read throughput in HA.
9. write benchmark -- Measures write throughput in HA.
The problem lies in the fact that there is still a record within the internal
storage of our old leader with the same transaction ID and GID as the recently
committed record by the new leader. Obviously, this is broken. As a solution, on
each transition from `Leader` to `Follower`, we will reinitialize storage, reset
the transaction engine and recover data from the Raft log. This will ensure all
ongoing transactions which have "polluted" the storage will be gone.
At the moment, our main goal is to pass existing tests and have a stable version
on our stress test. We should also implement a stress test which occasionally
introduces different types of failures in our cluster (we did this kind of
testing manually thus far). Passing these tests should convince us that we have
a "stable enough" version which we can start pushing to our customers.
"When will followers append that transaction to their commit logs?"
Additional (proper) testing should probably involve some ideas from
[here](https://jepsen.io/analyses/dgraph-1-0-2)
When the leader deduces that the transaction is safe to commit, it will include
the relevant information in all further heartbeats which will alert the
followers that it is safe to commit those entries from their raft logs.
Naturally, the followers need not to delay appending data to the commit log
as they know that the transaction has already been committed (from the clusters
point of view). If this sounds really messed up, seriously, read the Raft paper.
## Possible Future Changes/Improvements/Extensions
"How does the raft log differ from WAL"
There are two general directions in which we can alter HA Memgraph. The first
direction assumes we are going to stick with the Raft protocol. In that case
there are a few known ways to extend the basic algorithm in order to gain
better performance or achieve extra functionality. In no particular order,
these are:
Conceptually, it doesn't. When operating in HA, we don't really need the
recovery mechanisms implemented in Memgraph thus far. When a dead machine
comes back to life, it will eventually come in sync with the rest of the
cluster and everything will be done using the machine's raft log as well
as the messages received from the cluster leader.
1. Improving read performance using leader leases [Section 6.4 from Raft thesis]
2. Introducing cluster membership changes [Chapter 4 from Raft thesis]
3. Introducing a [learner mode](https://etcd.io/docs/v3.3.12/learning/learner/).
4. Consider different log compaction strategies [Chapter 5 from Raft thesis]
5. Removing HA proxy and implementing our own HA Memgraph client.
"Those logs will become huge, isn't that recovery going to be painfully slow?"
On the other hand, we might decide in the future to base our HA implementation
on a completely different protocol which might even offer different guarantees.
In that case we probably need to do a bit more of market research and weigh the
trade-offs of different solutions.
[This](https://www.postgresql.org/docs/9.5/different-replication-solutions.html)
might be a good starting point.
True, but there are mechanisms for making raft logs more compact. The most
popular method is, wait for it, making snapshots :)
Although the process of bringing an old machine back to life is a long one,
it doesn't really affect the performance of the cluster in a great degree.
The cluster will work perfectly fine with that machine being way out of sync.
## Reading materials
"I don't know, everything seems to be a lot slower than before!"
Absolutely true, the user should be aware that they will suffer dire
consequences on the performance side if they choose to be highly available.
As Frankie says, "That's life!".
"Also, I didn't really care about most of the things you've said. I'm
not a part of the storage team and couldn't care less about the issues you
face, how does HA affect 'my part of the codebase'?"
Answer for query execution: That's ok, you'll be able to use the same beloved
API (when we implement it, he he :) towards storage and continue to
make fun of us when you find a bug.
Answer for infrastructure: We'll talk. Some changes will surely need to
be made on the Memgraph client. There is a chapter in Diego's dissertation
called 'Client interaction', but we'll cross that bridge when we get there.
There will also be the whole 'integration with Jepsen tests' thing going on.
Answer for analytics: I'm astonished you've read this article. Wanna join
storage?
### Subtlety Regarding Reads
As we have hinted in the previous chapter, we would like to bypass log
replication for operations which do not alter the internal state of Memgraph.
Those operations should therefore be handled only by the leader, which is not
as trivial as it seems. The subtlety arises from the fact that a (newly-elected)
leader can have an entry in its log which was committed by the previous leader
that has crashed but that entry is not yet committed in its internal storage
by the current leader. Moreover, the rule about safely committing logs that are
replicated on the majority of the cluster only applies for entries replicated in
the leaders current term. Therefore, we are faced with two issues:
* We cannot simply perform read operations if the leader has a non-committed
entry in its log (breaks consistency).
* Replicating those entries onto the majority of the cluster is not enough
to guarantee that they can be safely committed.
This can be solved by introducing a blank no-op operation which the new leader
will try to replicate at the start of its term. Once that operation is
replicated and committed, the leader can safely perform those non-altering
operations on its own.
For further information about these issues, you should check out section
5.4.2 from the raft paper \[1\] which hints as to why its not safe to commit
entries from previous terms. Also, you should check out section 6.4 from
the thesis \[2\] which goes into more details around efficiently processing
read-only queries.
## How do we test HA
[Check this out](https://jepsen.io/analyses/dgraph-1-0-2)
1. [Raft paper](https://raft.github.io/raft.pdf)
2. [Raft thesis](https://github.com/ongardie/dissertation) (book.pdf)
3. [Raft playground](https://raft.github.io/)
4. [Leader Leases](https://blog.yugabyte.com/low-latency-reads-in-geo-distributed-sql-with-raft-leader-leases/)
5. [Improving Raft ETH](https://pub.tik.ee.ethz.ch/students/2017-FS/SA-2017-80.pdf)