Comments below..
On 5/21/2017 1:38 PM, Monty Taylor wrote:
Hi all!
As the discussion around PostgreSQL has progressed, it has come clear
to me that there is a decently deep philosophical question on which we
do not currently share either definition or agreement. I believe that
the lack of clarity on this point is one of the things that makes the
PostgreSQL conversation difficult.
I believe the question is between these two things:
* Should OpenStack assume the existence of an external database
service that it treat as an black-box on the other side of a
connection string?
* Should OpenStack take an active and/or opinionated role in managing
the database service?
A potentially obvious question about that (asked by Mike Bayer in a
different thread) is: "what do you mean by managing?"
What I mean by managing is doing all of the things you can do related
to database operational controls short of installing the software,
writing the basic db config files to disk and stopping and starting
the services. It means being much more prescriptive about what types
of config we support, validating config settings that cannot be
overridden at runtime and refusing to operate if they are unworkable.
I think it's helpful and important for us to have automation tooling
like tripleo, puppet, etc. that can stand up a MySQL database. But we
also have to realize that as shops mature, they will deploy more
complicated database topologies, clustered configurations, and
replication scenarios. So I think we shouldn't go overboard with being
prescriptive. We also have to realize that in the enterprise space,
databases are usually deployed and managed by a separate database team,
which means less control over that layer. So we shouldn't force people
into this model. We should provide best practice documentation, examples
(tripleo, puppet, ansible, etc.), and leave it up to the operator.
Why would we want to be 'more active'? When managing and tuning
databases, there are some things that are driven by the environment
and some things that are driven by the application.
Things that are driven by the environment include things like the
amount of RAM actually available, whether or not the machines running
the database are dedicated or shared, firewall settings, selinux
settings and what versions of software are available.
This is a good example of an area that we should focus on documenting
best practices and leave it to the operator to implement. Guidelines
around cpu, memory, security settings, tunables, etc. are what's needed
here. Today, there isn't really any guidance or best practices on even
sizing the database(s) for a given deployment size.
Things that are driven by the application are things like character
set and collation, schema design, data types, schema upgrade and HA
strategies.
These are things that we can have a bit more control or direction on.
One might argue that HA strategies are an operator concern, but in
reality the set of workable HA strategies is tightly constrained by
how the application works, and the pairing an application expecting
one HA strategy with a deployment implementing a different one can
have negative results ranging from unexpected downtime to data
corruption.
For example: An HA strategy using slave promotion and a VIP that
points at the current write master paired with an application
incorrectly configured to do such a thing can lead to writes to the
wrong host after a failover event and an application that seems to be
running fine until the data turns up weird after a while.
This is definitely a more complicated area that becomes more and more
specific to the clustering technology being used. Galera vs. MySQL
Cluster is a good example. Galera has an active/passive architecture
where the above issues become a concern for sure. While MySQL Cluster
(NDB) is an active/active architecture, so losing a node only effects
any uncommitted transactions, that could easily be addressed with a
retry. These topologies will become more complicated as people start
looking at cross regional replication and DR.
For the areas in which the characteristics of the database are tied
closely to the application behavior, there is a constrained set of
valid choices at the database level. Sometimes that constrained set
only has one member.
The approach to those is what I'm talking about when I ask the
question about "external" or "active".
In the "external" approach, we document the expectations and then
write the code assuming that the database is set up appropriately. We
may provide some helper tools, such as 'nova-manage db sync' and
documentation on the sequence of steps the operator should take.
In the "active" approach, we still document expectations, but we also
validate them. If they are not what we expect but can be changed at
runtime, we change them overriding conflicting environmental config,
and if we can't, we hard-stop indicating an unsuitable environment.
Rather than providing helper tools, we perform the steps needed
ourselves, in the order they need to be performed, ensuring that they
are done in the manner in which they need to be done.
This might be a trickier situation, especially if the database(s) are in
a separate or dedicated environment that the OpenStack service processes
don't have access to. Of course for SQL commands, this isn't a problem.
But changing the configuration files and restarting the database may be
a harder thing to expect.
Some examples:
* Character Sets / Collations
We currently enforce at testing time that all database migrations are
explicit about InnoDB. We also validate in oslo.db that table
character sets have the string 'utf8' in them. (only on MySQL) We do
not have any check for case-sensitive or case-insensitive collations
(these affect sorting and comparison operations) Because we don't,
different server config settings or different database backends for
different clouds can actually behave differently through the REST API.
To deal with that:
First we'd have to decide whether case sensitive or case insensitive
was what we wanted. If we decided we wanted case sensitive, we could
add an enforcement of that in oslo.db, and write migrations to get
from case insensitive indexes to case sensitive indexes on tables
where we detected that a case insensitive collation had been used. If
we decided we wanted to stick with case insensitive we could similarly
add code to enforce it on MySQL. To enforce it actively on
PostgresSQL, we'd need to either switch our code that's using
comparisons to use the sqlalchemy case-insensitive versions
explicitly, or maybe write some sort of overloaded driver for PG that
turns all comparisons into case-insensitive, which would wrap both
sides of comparisons in lower() calls (which has some indexing
concerns, but let's ignore that for the moment) We could also take the
'external' approach and just document it, then define API tests and
try to tie the insensitive behavior in the API to Interop Compliance.
I'm not 100% sure how a db operator would remediate this - but PG has
some fancy computed index features - so maybe it would be possible.
I think that abstraction with oslo.db would be the right path here. But
you are also right that we need to have a consistent compliance policy
at the API layer. We may fix things down at the DB level with oslo.db,
but everything on top of that needs to also fall in-line. There is a
very high chance that there are hard-coded workarounds or assumptions in
the services and apis today.
A similar issue lurks with the fact that MySQL unicode storage is
3-byte by default and 4-byte is opt-in. We could take the 'external'
approach and document it and assume the operator has configured their
my.cnf with the appropriate default, or taken an 'active' approach
where we override it in all the models and make migrations to get us
from 3 to 4 byte.
I think an active approach on this would be ideal, just like the utf8
and InnoDB settings are today. FYI, not all services are enforcing these
in a consistent manor today. Another example of something that should be
abstracted at the oslo.db layer and get the human element out of the way.
* Schema Upgrades
The way you roll out online schema changes is highly dependent on your
database architecture.
Just limiting to the MySQL world:
If you do Galera, you can do roll them out in Total Order or Rolling
fashion. Total Order locks basically everything while it's happening,
so isn't a candidate for "online". In rolling you apply the schema
change to one node at a time. If you do that, the application has to
be able to deal with both forms of the table, and you have to deal
with ensuring that data can replicate appropriately while the schema
change is happening.
If you do DRBD active/passive or a single-node deployment you only
have one upgrade operation to perform, but you will only lock certain
things - depending on what schema change operations you were performing.
If you do master/slave, you can roll out the schema change to your
slaves one at a time, wait for them all to catch up, then promote a
slave taking the current master out of commission - update the old
master then then put it into the slave pool. Like Galera rolling, the
app needs to be able to handle old and new versions and the
replication stream needs to be able to replicate between the versions.
Making sure that the stream is able to replicate puts a set of
limitations on the types of schema changes you can perform, but it is
an understandable constrained set.
In either approach the OpenStack service has to be able to talk to
both old and new versions of the schema. And in either approach we
need to make sure to limit the schema change operations to the set
that can be accomplished in an online fashion. We also have to be
careful to not start writing values to new columns until all of the
nodes have been updated, because the replication stream can't
replicate the new column value to nodes that don't have the new column.
This is another area where something like MySQL Cluster (NDB) would
operate differently because it's an active/active architecture. So
limiting the number of online changes while a table is locked across the
cluster would be very important. There is also the timeouts for the
applications to consider, something that could be abstracted again with
oslo.db.
In either approach we can decide to limit the number of architectures
we support for "online" upgrades.
In an 'external' approach, we make sure to do those things, we write
documentation and we assume the database will be updated
appropriately. We can document that if the deployer chooses to do
Total Order on Galera, they will not have online upgrades. There will
also have to be a deployer step to let the services know that they can
start writing values to the new schema format once the upgrade is
complete.
In an 'active' approach, we can notice that we have an update
available to run, and we can drive it from code. We can check for
Galera, and if it's there we can run the upgrade in Rolling fashion
one node at a time with no work needed on the part of the deployer.
Since we're driving the upgrade, we know when it's done, so we can
signal ourselves to start using the new version. We'd obviously have
to pick the set of acceptable architectures we can handle consistently
orchestrating.
This would be an interesting idea to expand to a autonomic orchestration
framework within the control plane to handle the database upgrades
online and the restarting of the dependent services in the correct
order. If we only focus on the database piece, it may not be as
interesting for operators.
* Versions
It's worth noting that behavior for schema updates and other things
change over time with backend database version. We set minimum
versions of other things, like libvirt and OVS - so we might also want
to set minimum versions for what we can support in the database. That
way we can know for a given release of OpenStack what DDL operations
are safe to use for a rolling upgrade and what are not. That means
detecting such a version and potentially refusing to perform an
upgrade if the version isn't acceptable. That reduces the operator's
ability to choose what version of the database software to run, but
increases our ability to be able to provide tooling and operations
that we can be confident will work.
Validating the MySQL database version is a good idea. The features do
change over time. A good example is how in 5.7, you'll get warnings
about duplicate indexes being dropped in a future release which will
definitely affect multiple services today.
== Summary ==
These are just a couple of examples - but I hope they're at least
mildly useful to explain some of the sorts of issues at hand - and why
I think we need to clarify what our intent is separate from the issue
of what databases we "support".
Some operations have one and only one "right" way to be done. For
those operations if we take an 'active' approach, we can implement
them once and not make all of our deployers and distributors each
implement and run them. However, there is a cost to that. Automatic
and prescriptive behavior has a higher dev cost that is proportional
to the number of supported architectures. This then implies a need to
limit deployer architecture choices.
On the other hand, taking an 'external' approach allows us to federate
the work of supporting the different architectures to the deployers.
This means more work on the deployer's part, but also potentially a
greater amount of freedom on their part to deploy supporting services
the way they want. It means that some of the things that have been
requested of us - such as easier operation and an increase in the
number of things that can be upgraded with no-downtime - might become
prohibitively costly for us to implement.
I honestly think that both are acceptable choices we can make and that
for any given topic there are middle grounds to be found at any given
moment in time.
BUT - without a decision as to what our long-term philosophical intent
in this space is that is clear and understandable to everyone, we
cannot have successful discussions about the impact of implementation
choices, since we will not have a shared understanding of the problem
space or the solutions we're talking about.
For my part - I hear complaints that OpenStack is 'difficult' to
operate and requests for us to make it easier. This is why I have been
advocating some actions that are clearly rooted in an 'active' worldview.
Finally, this is focused on the database layer but similar questions
arise in other places. What is our philosophy on prescriptive/active
choices on our part coupled with automated action and ease of
operation vs. expanded choices for the deployer at the expense of
configuration and operational complexity. For now let's see if we can
answer it for databases, and see where that gets us.
Thanks for reading.
Monty
__________________________________________________________________________
OpenStack Development Mailing List (not for usage questions)
Unsubscribe:
openstack-dev-requ...@lists.openstack.org?subject:unsubscribe
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
__________________________________________________________________________
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev