Previously, if one of these tests failed to start their servers before
the timeout, there was a risk that the goroutine started to run
the server would invoke t.Errorf after the test itself had concluded.
Replace t.Errorf with fmt.Printf to avoid the panic, and include
the test name for debugging.
* internal/handlers/managed_groups: ensure membership lookup is unlimited
Previously, we would only show the first 10,000 members in the group,
giving the impression that a complete list was performed. Now we
list all members.
* internal/handlers/accounts: ensure group lookup is unlimited
Previously, we would only show the first 10,000 managed group
memberships. Now we list all group memberships of an account.
Previously, we would list the managed groups with an implied
filter of db.DefaultLimit (10,000), which would incorrectly remove
a user from managed group memberships if there were more
than 10,000 managed groups in an auth method. Explicitly
set the list limit to unlimited to ensure all managed groups are
updated appropriately.
This commit changes the logic around selecting jobs to run. Before this,
Boundary used a view "job_jobs_to_run" which was a collation of all jobs
that were not running and scheduled to run. When jobs were requested to
run, Boundary would look at this view, grab a certain amount of jobs and
insert them into the "job_run" table. Importantly, while the amount of
jobs to be scheduled at once defaulted to 1 in the scheduler package,
this was entirely overriden in the controller package to -1 (no limit,
run all jobs that the query would return).
With this commit, Boundary now looks at the "job" table directly for
jobs to run, regardless of job run status. This was done to improve DB
performance as the view (which queried all "job_run" rows) was causing
significant DB contention.
Additionally, the option to limit job runs is now removed since it was
already hardcoded to -1 and not user-configurable anyways.
Should the need arise to implement this limit again, this will have to
be done in a different way since the new query does not account for
running jobs and limiting the rows returned after jobs are already
running would yield a no-op (the new query inserts nothing into
"job_run" if any given job is already running).
Before this change, upon the completion of a job run, the scheduler
would update a job_run entry with the completed status as well as the
number of objects affected by the run. Additionally, we then had a
cleaner job whose scope was deleting these completed entries from
job_run.
This commit changes this design to drop the cleaner job the and rolls
its functionality directly into the scheduler. Now, when any given job
run is complete, the scheduler will automatically delete it from the
job_run table itself.
In addition, the completed status has been removed as a valid status as
it is not set anymore.
This change follows a decision that given the data in these job_run
entries is already being deleted by the cleaner job, updating its state
when a job completes just to have it deleted is not efficient.
Additionally, no use-case for this data was surfaced.
This changes the periodic job that deletes terminated sessions to delete
the those sessions in batches. Prior to this, a single SQL DELETE
command was used to delete all qualifying terminated sessions. In
systems under heavy load, the statement could timeout leading to a
situation where no qualifying terminated sessions were ever being
deleted.
This adds retries as an attribute to JobStatus allowing a job to
indicate liveness to the scheduler when work has not progressed.
A job could use this and the statusThreshold passed into the Run method
to break work up into batches that would otherwise timeout if all the
work was attempted at once, for example a long running SQL delete
command.
The scheduler monitors running jobs to make sure they are not hung. If
the status of a running job does not change after a duration (the
interruptThreshold in the scheduler), the scheduler kills the job. This
change passes the scheduler's interruptThreshold to the job when it
calls Run() on the job.
When a resource relating to a session is deleted,
notably a target, host, or auth_token, the corresponding sessions are
marked as canceled via a database trigger that calls the cancel_session
function. This function was performing a select as part of the insert
call to ensure that it would not attempt to insert a session_sate of
canceled if it was an invalid state transition. However, since this
function was first introduced, there has been additional constraints
added that would prevent an invalid state transition. In addition,
performing a select as part of an insert increases the chances for
contention with other queries that interact with the session_state
table, especially since a large number of sessions and session_state
rows can be included in a single transaction when things like
auth_tokens or targets get deleted.
Now, the cancel_function will perform the insert, and catch any of the
potential violations, allowing the query to continue without needing to
perform the select, thus reduction the potential for contention.
This issue was surfaced through other changes (PR 5137).
Given that the eventer works in a singleton pattern (see sysEventer
global variable), performing this shallow object copy can cause a data
race because the copy operation reads the underlying eventer data
structures. At the same time, given the global nature of this eventer,
other goroutines can be writing to those same data structures.
Since we're already using a singleton pattern in this package, and given
that the current copy operation is shallow, copying the object doesn't
seem necessary, so this commit changes `StandardWriter` to build the
`logAdapter` with the `e` eventer directly instead of copying it.
This replaces the start_time, end_time, and previous_end_time columns on
the session_state table with a time range column that records when the
state was active. This allowed for dropping several constraints on the
session_state table that were used to ensure that states for a session
did not overlap. Instead this can be ensured with a single exclusion
constraint. By reducing the number of constraints, it improves the
performance for a number of queries involved with session state. Most
notably when resolving delete cascades of resources related to sessions,
such as targets, where as part of the delete transaction large numbers
of sessions get canceled. Similarly, when sessions get deleted, and the
corresponding session_state rows are deleted, these constraints all must
be checked during the transaction.
The domain layer continues to expose StartTime and EndTime fields on
session.State. However it drops PreviousEndTime, which was never used by
anything calling the domain layer.
See:
https://www.postgresql.org/docs/current/ddl-constraints.html#DDL-CONSTRAINTS-EXCLUSIONhttps://www.postgresql.org/docs/current/rangetypes.html
This function was returning the set of deletion tables. A view seems
better suited for this task since it would allow for applying additional
filters of the result set. This was particularly necessary to easily
make changes to some sqltests due to switching the delete trigger for
the session table.
(cherry picked from commit 32d41639892950b21fafc39b50c10b997c6d5121)
When sessions are deleted, a trigger is used to insert records into the
session_deleted table. This table is utilized by the sessions list
endpoint when using a refresh token to inform a client of the sessions
that have been deleted since the last request. We delete sessions in
bulk via a controller job to delete sessions that have been terminated
over an hour ago, which results in the trigger running a large number of
separate insert statements while processing the delete statement.
This changes the trigger to run once for the delete statement, instead
of for each row, resulting in a single bulk insert statement to the
session_deleted table.
This new trigger function also avoids the use of `on conflict`. When
testing this function, while the single statement was still faster than
running multiple inserts, the `on conflict` still added significant
overhead, even when there were no conflicts. It should be safe to
perform the insert without the `on conflict`, since the same ID should
never be deleted more than once if it is successfully deleted.
(cherry picked from commit 3569b36c6eb2bee7a4e36e7584fcc6bbf1124f95)
When processing a controller API request, system uses a query to fetch
the grants for the requesting user. This query requires pulling together
information from several tables, and is currently performing many
sequential scans to do so. For a number of the tables involved, there
are indexes from multi-column primary keys, however, due to the order of
the columns in the index, the postgres planner would need to do a full
scan of the index, which can be less efficient than a sequential scan,
so instead it will perform a sequential scan of the table.
This recreates the primary keys while swapping the order of the columns
in the primary key definition, and thus the order in the index. By doing
so, the planner will not need to perform a full index scan, and will be
more likely to use the index when executing the grants query.
(cherry picked from commit 5577b3c358e9d45cd7ec02fc951e45d52e50595e)
This adds indexes for a few categories:
1. Foreign keys on the session table. These are set to `null` when the
referenced row is deleted. These indexes will help to more
efficiently set these `null` values in the case where a target is
deleted, or and auth token is deleted.
2. Foreign keys from other tables to the session table. These are either
set to `null` or cascade deleted when a session is deleted. These
indexes help with these update/deletes when a session is deleted.
3. A multi-column index on session_state. This helps with the query that
is used to delete all terminated sessions that have been terminated
for over an hour.
(cherry picked from commit 263a96e9ddd197cb36f2c4c74e91ac034112321e)
This removes the Cartesian product in the DB for GrantsForUser in favor
of returning the actual grant scope information and dealing with it in
application code.
(cherry picked from commit 913cd536823d73994acc4f4d833208c1cc05f91b)
* internal/clientcache: add -force-reset-schema flag
* clientcache: stream list pages directly to DB
---------
Co-authored-by: Johan Brandhorst-Satzkorn <johan.brandhorst@gmail.com>
(cherry picked from commit ccea8a4ce53f7b0d1105c1c1e4875e443635d96a)
Fix ensures that the TestServer returned uses the
correct directory for writing its sqlite files,
which was created via t.TempDir()
(cherry picked from commit 85f91043a164bdacc1c6738ae0e7c50077d55c79)
Johan Brandhorst-Satzkorn
Michael Gaffney
Elim Tsiagbey
Hugo
daniandDamian Debkowski
Irena Rindos
Jim
Cameron Perera
Lisbet Alvarez
Timothy Messier
Ryan Derr
Jeff Mitchell