A static initialization is not desirable in the main binaries, as it forces all
paths of code to init, but it is still useful in tests. It allows static
intialization to be performed once for all tests and eliminates the need to
always add the FluxInit call. Added a fluxinit/static package that calls
fluxinit.FluxInit() to replace the builtin package. This hides the nature of
the initialization and makes it clear that it is mandatory initialization code
getting called.
The controller now supports setting an initial memory limit and setting
a maximum amount of memory that the controller may use separately from
the memory quota per query and the concurrency quota.
This allows the controller to increase the concurrency quota to a larger
number while setting the maximum amount of memory to a lower amount than
would be required for all queries to use 100% of their allowable memory.
Functionally, this means that a query will have a soft limit for an
initial memory byte quota that a query is guaranteed to have, a shared
pool that it is allowed access to in the case it uses more, and a hard
limit that no query may exceed to prevent runaway queries from taking
over the entire pool.
This change is completely backwards compatible with older configurations
as the new options will default to values that mimic the old behavior
where a query is allocated the full amount of its memory quota and the
maximum amount of memory is based on the concurrency quota and this
maximum memory quota.
In addition to the above, this also fixes a bug in the controller that
allowed it to run more than its concurrency as executing queries. This
happened when the results had finished being sent by the executor, but
the query had not yet been read and/or serialized. The executor would be
freed up and would take the next query even though the previous query
hadn't yet been finalized with `Done()`.
The http error schema has been changed to simplify the outward facing
API. The `op` and `error` attributes have been dropped because they
confused people. The `error` attribute will likely be readded in some
form in the future, but only as additional context and will not be
required or even suggested for the UI to use.
Errors are now output differently both when they are serialized to JSON
and when they are output as strings. The `op` is no longer used if it is
present. It will only appear as an optional attribute if at all. The
`message` attribute for an error is always output and it will be the
prefix for any nested error. When this is serialized to JSON, the
message is automatically flattened so a nested error such as:
influxdb.Error{
Msg: errors.New("something bad happened"),
Err: io.EOF,
}
This would be written to the message as:
something bad happened: EOF
This matches a developers expectations much more easily as most
programmers assume that wrapping an error will act as a prefix for the
inner error.
This is flattened when written out to HTTP in order to make this logic
immaterial to a frontend developer.
The code is still present and plays an important role in categorizing
the error type. On the other hand, the code will not be output as part
of the message as it commonly plays a redundant and confusing role when
humans read it. The human readable message usually gives more context
and a message like with the code acting as a prefix is generally not
desired. But, the code plays a very important role in helping to
identify categories of errors and so it is very important as part of the
return response.
We are planning to change the allocator interface within flux to use the
arrow allocator. To make the release easier, this updates the test in
advance to use the arrow allocator instead of the to be changed memory
allocator interface from flux.
If we handle the flux errors in the query controller, it makes it so we
are handling the errors in the location where the happen rather than at
a layer further up the stack.
This should simplify it so the errors are handled in this single
location instead.
In the QueueSize test, it was possible that after the `done` channel was
closed, one of the queries from the queue would begin executing. If all
three began executing before the shutdown was done, the third would
block on attempting to send a value to the `executing` channel and it
would never finish so the controller would report that shutdown failed.
This increases the queue size to a combination of the concurrency quota
and the queue size so all of the started queries will never block when
sending a signal to the executing channel.
The controller implementation is primarily used by influxdb so it
shouldn't be part of the flux repository. This copies the code from flux
to influxdb so it can be removed from the next flux release.
Daniel Moran
Adrian Thurston
Christopher M. Wolff
Jonathan A. Sternberg
Lorenzo Affetti
Nathaniel Cook