nova/doc/source/cells.rst
Matt Riedemann 409085a8c0 docs: link to Laski's cells talk from the Austin summit
Andrew Laski gave a nice presentation at the Austin Newton summit
of cells v1 and v2, along with work complete and what's being
worked on. This fits nicely into our documentation on cells and is
good to watch before reading further in detail.

Change-Id: I4c0245c000bb74c159fc02c36a83a11f145208a3
2016-05-02 12:23:08 -04:00

8.5 KiB

Cells

Before reading further, there is a nice overview presentation that Andrew Laski gave at the Austin (Newton) summit which is worth watching.

Cells V1

Historically, Nova has depended on a single logical database and message queue that all nodes depend on for communication and data persistence. This becomes an issue for deployers as scaling and providing fault tolerance for these systems is difficult.

We have an experimental feature in Nova called "cells", hereafter referred to as "cells v1", which is used by some large deployments to partition compute nodes into smaller groups, coupled with a database and queue. This seems to be a well-liked and easy-to-understand arrangement of resources, but the implementation of it has issues for maintenance and correctness. See Comparison with Cells V1 for more detail.

Status

Cells v1 is considered experimental and receives much less testing than the rest of Nova. For example, there is no job for testing cells v1 with Neutron.

The priority for the core team is implementation of and migration to cells v2. Because of this, there are a few restrictions placed on cells v1:

  1. Cells v1 is in feature freeze. This means no new feature proposals for cells v1 will be accepted by the core team, which includes but is not limited to API parity, e.g. supporting virtual interface attach/detach with Neutron.
  2. Latent bugs caused by the cells v1 design will not be fixed, e.g. bug 1489581. So if new tests are added to Tempest which trigger a latent bug in cells v1 it may not be fixed. However, regressions in working function should be tracked with bugs and fixed.

Suffice it to say, new deployments of cells v1 are not encouraged.

The restrictions above are basically meant to prioritize effort and focus on getting cells v2 completed, and feature requests and hard to fix latent bugs detract from that effort. Further discussion on this can be found in the 2015/11/12 Nova meeting minutes.

There are no plans to remove Cells V1 until V2 is usable by existing deployments and there is a migration path.

Cells V2

Manifesto

Proposal

Right now, when a request hits the Nova API for a particular instance, the instance information is fetched from the database, which contains the hostname of the compute node on which the instance currently lives. If the request needs to take action on the instance (which is most of them), the hostname is used to calculate the name of a queue, and a message is written there which finds its way to the proper compute node.

The meat of this proposal is changing the above hostname lookup into two parts that yield three pieces of information instead of one. Basically, instead of merely looking up the name of the compute node on which an instance lives, we will also obtain database and queue connection information. Thus, when asked to take action on instance $foo, we will:

  1. Lookup the three-tuple of (database, queue, hostname) for that instance
  2. Connect to that database and fetch the instance record
  3. Connect to the queue and send the message to the proper hostname queue

The above differs from the current organization in two ways. First, we need to do two database lookups before we know where the instance lives. Second, we need to demand-connect to the appropriate database and queue. Both of these have performance implications, but we believe we can mitigate the impacts through the use of things like a memcache of instance mapping information and pooling of connections to database and queue systems. The number of cells will always be much smaller than the number of instances.

There are availability implications with this change since something like a 'nova list' which might query multiple cells could end up with a partial result if there is a database failure in a cell. A database failure within a cell would cause larger issues than a partial list result so the expectation is that it would be addressed quickly and cellsv2 will handle it by indicating in the response that the data may not be complete.

Since this is very similar to what we have with current cells, in terms of organization of resources, we have decided to call this "cellsv2" for disambiguation.

After this work is complete there will no longer be a "no cells" deployment. The default installation of Nova will be a single cell setup.

Benefits

The benefits of this new organization are:

  • Native sharding of the database and queue as a first-class-feature in nova. All of the code paths will go through the lookup procedure and thus we won't have the same feature parity issues as we do with current cells.
  • No high-level replication of all the cell databases at the top. The API will need a database of its own for things like the instance index, but it will not need to replicate all the data at the top level.
  • It draws a clear line between global and local data elements. Things like flavors and keypairs are clearly global concepts that need only live at the top level. Providing this separation allows compute nodes to become even more stateless and insulated from things like deleted/changed global data.
  • Existing non-cells users will suddenly gain the ability to spawn a new "cell" from their existing deployment without changing their architecture. Simply adding information about the new database and queue systems to the new index will allow them to consume those resources.
  • Existing cells users will need to fill out the cells mapping index, shutdown their existing cells synchronization service, and ultimately clean up their top level database. However, since the high-level organization is not substantially different, they will not have to re-architect their systems to move to cellsv2.
  • Adding new sets of hosts as a new "cell" allows them to be plugged into a deployment and tested before allowing builds to be scheduled to them.

Comparison with Cells V1

In reality, the proposed organization is nearly the same as what we currently have in cells today. A cell mostly consists of a database, queue, and set of compute nodes. The primary difference is that current cells require a nova-cells service that synchronizes information up and down from the top level to the child cell. Additionally, there are alternate code paths in compute/api.py which handle routing messages to cells instead of directly down to a compute host. Both of these differences are relevant to why we have a hard time achieving feature and test parity with regular nova (because many things take an alternate path with cells) and why it's hard to understand what is going on (all the extra synchronization of data). The new proposed cellsv2 organization avoids both of these problems by letting things live where they should, teaching nova to natively find the right db, queue, and compute node to handle a given request.

Database split

As mentioned above there is a split between global data and data that is local to a cell.

The following is a breakdown of what data can uncontroversially considered global versus local to a cell. Missing data will be filled in as consensus is reached on the data that is more difficult to cleanly place. The missing data is mostly concerned with scheduling and networking.

Global (API-level) Tables

instance_types instance_type_projects instance_type_extra_specs quotas project_user_quotas quota_classes quota_usages security_groups security_group_rules security_group_default_rules provider_fw_rules key_pairs migrations networks tags

Cell-level Tables

instances instance_info_caches instance_extra instance_metadata instance_system_metadata instance_faults instance_actions instance_actions_events instance_id_mappings pci_devices block_device_mapping virtual_interfaces