Database Reference
In-Depth Information
Applicationhost Runs SPL jobs.
Managementhost Runs the management services that control the flow
of SPL jobs (but doesn't explicitly run any SPL jobs directly), manages
security within the instance, monitors any running jobs, and so on.
Mixedhost
Can run both SPL jobs and management tasks.
In a typical small environment, you would have one management host and
the remainder of your servers would be used as application hosts, but for
large deployments, management services can be distributed to more hosts.
When you build a streaming application, operators are compiled into one
or more processing elements (PEs). PEs can contain one or more operators,
which are often “fused” together inside a single PE for performance reasons.
A PE can be considered to be a unit of physical deployment in the Streams
product. PEs from the same application can run on multiple hosts in a
network, as well as communicate tuples across the network. In the event of a
PE failure, Streams automatically detects the failure and chooses from among
a large number of possible remedial actions. For example, if the PE is restart-
able and relocatable, the Streams run-time engine automatically picks an
available host on which to run the PE; starts the PE on that host; and auto-
matically “rewires” inputs and outputs to other PEs, as appropriate. How-
ever, if the PE continues to fail over and over again and exceeds a retry threshold
(perhaps due to a recurring underlying hardware issue), the PE is placed into a
stopped state and requires manual intervention to resolve the issue. If the
PE is restartable, but has been defined as not relocatable (for example, the PE
is a sink that requires it to be run on a specific host), the Streams run-time
engine automatically attempts to restart the PE on the same host, if it is avail-
able. Similarly, if a management host fails, you can have the management
function restarted elsewhere, assuming that you have configured the system
with RecoveryMode=ON . In this case, the recovery metadata has the neces-
sary information to restart the management tasks on another server in the
cluster.
Data processing is guaranteed in Streams when there are no host,
network, or PE failures. You might be wondering what happens to the data
in the event of one of these failures? When a PE (or its network or host) does
fail, the data in the PE's buffers (and data that appears while a PE is being
restarted) can be lost without special precautions. With the most sensitive
Search WWH ::




Custom Search