Query nodes table using AS OF SYSTEM TIME '-10s' (by default) when on CRDB to alleviate contention on the nodes table and minimize CRDB retries. Queries for standard uploads are already cached, and node lookups for graceful exit uploads has retry logic so it isn't necessary for the nodes returned to be current.
The chief segment health models we've come up with are the "immediate
danger" model and the "survivability" model. The former calculates the
chance of losing a segment becoming lost in the next time period (using
the CDF of the binomial distribution to estimate the chance of x nodes
failing in that period), while the latter estimates the number of
iterations for which a segment can be expected to survive (using the
mean of the negative binomial distribution). The immediate danger model
was a promising one for comparing segment health across segments with
different RS parameters, as it is more precisely what we want to
prevent, but it turns out that practically all segments in production
have infinite health, as the chance of losing segments with any
reasonable estimate of node failure rate is smaller than DBL_EPSILON,
the smallest possible difference from 1.0 representable in a float64
(about 1e-16).
Leaving aside the wisdom of worrying about the repair of segments that
have less than a 1e-16 chance of being lost, we want to be extremely
conservative and proactive in our repair efforts, and the health of the
segments we have been repairing thus far also evaluates to infinity
under the immediate danger model. Thus, we find ourselves reaching for
an alternative.
Dr. Ben saves the day: the survivability model is a reasonably close
approximation of the immediate danger model, and even better, it is
far simpler to calculate and yields manageable values for real-world
segments. The downside to it is that it requires as input an estimate
of the total number of active nodes.
This change replaces the segment health calculation to use the
survivability model, and reinstates the call to SegmentHealth() where it
was reverted. It gets estimates for the total number of active nodes by
leveraging the reliability cache.
Change-Id: Ia5d9b9031b9f6cf0fa7b9005a7011609415527dc
A few weeks ago it was discovered that the segment health function
was not working as expected with production values. As a bandaid,
we decided to insert the number of healthy pieces into the segment
health column. This should have effectively reverted our means of
prioritizing repair to the previous implementation.
However, it turns out that the bandaid was placed into the code which
removes items from the irreparable db and inserts them into the repair
queue.
This change: insert number of healthy pieces into the repair queue in the
method, RemoteSegment
Change-Id: Iabfc7984df0a928066b69e9aecb6f615253f1ad2
There is a new checker field called statsCollector. This contains
a map of stats pointers where the key is a stringified redundancy
scheme. stats contains all tagged monkit metrics. These metrics exist
under the key name, "tagged_repair_stats", which is tagged with the
name of each metric and a corresponding rs scheme.
As the metainfo observer works on a segment, it checks statsCollector
for a stats corresponding to the segment's redundancy scheme. If one
doesn't exist, it is created and chained to the monkit scope. Now we can call
Observe, Inc, etc on the fields just like before, and they have tags!
durabilityStats has also been renamed to aggregateStats.
At the end of the metainfo loop, we insert the aggregateStats totals into the
corresponding stats fields for metric reporting.
Change-Id: I8aa1918351d246a8ef818b9712ed4cb39d1ea9c6
We migrated satelliteDB off of Postgres and over to CockroachDB (crdb), but there was way too high contention for the injuredsegments table so we had to rollback to Postgres for the repair queue. A couple things contributed to this problem:
1) crdb doesn't support `FOR UPDATE SKIP LOCKED`
2) the original crdb Select query was doing 2 full table scans and not using any indexes
3) the SLC Satellite (where we were doing the migration) was running 48 repair worker processes, each of which run up to 5 goroutines which all are trying to select out of the repair queue and this was causing a ton of contention.
The changes in this PR should help to reduce that contention and improve performance on CRDB.
The changes include:
1) Use an update/set query instead of select/update to capitalize on the new `UPDATE` implicit row locking ability in CRDB.
- Details: As of CRDB v20.2.2, there is implicit row locking with update/set queries (contention reduction and performance gains are described in this blog post: https://www.cockroachlabs.com/blog/when-and-why-to-use-select-for-update-in-cockroachdb/).
2) Remove the `ORDER BY` clause since this was causing a full table scan and also prevented the use of the row locking capability.
- While long term it is very important to `ORDER BY segment_health`, the change here is only suppose to be a temporary bandaid to get us migrated over to CRDB quickly. Since segment_health has been set to infinity for some time now (re: https://review.dev.storj.io/c/storj/storj/+/3224), it seems like it might be ok to continue not making use of this for the short term. However, long term this needs to be fixed with a redesign of the repair workers, possible in the trusted delegated repair design (https://review.dev.storj.io/c/storj/storj/+/2602) or something similar to what is recommended here on how to implement a queue on CRDB https://dev.to/ajwerner/quick-and-easy-exactly-once-distributed-work-queues-using-serializable-transactions-jdp, or migrate to rabbit MQ priority queue or something similar..
This PRs improved query uses the index to avoid full scans and also locks the row its going to update and CRDB retries for us if there are any lock errors.
Change-Id: Id29faad2186627872fbeb0f31536c4f55f860f23
the immediate need is to be able to move the repair queue back out
of cockroach if we can't save it.
Change-Id: If26001a4e6804f6bb8713b4aee7e4fd6254dc326
We did not test the SegmentHealth function with actual production
values, and it turns out that values such as 52 healthy, 35 minimum
result in +Inf segment health - so pretty much all segments put into the
repair queue have the same health, which means we effectively aren't
sorting by health.
This change inserts numHealthy as segment health into the database so
the segments are ordered as they were before. We need to refine the
SegmentHealth function before we can support multi RS.
Change-Id: Ief19bbfee3594c5dfe94ca606bc930f05f85ff74
Rather than having a single repair override value, we will now support
repair override values based on a particular segment's RS scheme.
The new format for RS override values is
"k/o/n-override,k/o/n-override..."
Change-Id: Ieb422638446ef3a9357d59b2d279ee941367604d
Firstly, this changes the repair functionality to return Canceled errors
when a repair is canceled during the Get phase. Previously, because we
do not track individual errors per piece, this would just show up as a
failure to download enough pieces to repair the segment, which would
cause the segment to be added to the IrreparableDB, which is entirely
unhelpful.
Then, ignore Canceled errors in the return value of the repair worker.
Apparently, when the worker returns an error, that makes Cobra exit the
program with a nonzero exit code, which causes some piece of our
deployment automation to freak out and page people. And when we ask the
repair worker to shut down, "canceled" errors are what we _expect_, not
an error case.
Change-Id: Ia3eb1c60a8d6ec5d09e7cef55dea523be28e8435
We plan to add support for a new Reed-Solomon scheme soon, but our
repair queue orders segments by least number of healthy pieces first.
With a second RS scheme, fewer healthy pieces will not necessarily
correlate to lower health.
This change just adds the new column in a migration. A separate change
will add the new health function.
Right now, since we only support one RS scheme, behavior will not
change. Number of healthy pieces is being inserted as "segment health"
until the new health function is merged.
Segment health is calculated with a new priority function created in
commit 3e5640359. In order to use the function, a new config value is
added, called NodeFailureRate, representing the approximate probability
of any individual node going down in the duration of one checker run.
Change-Id: I51c4202203faf52528d923befbe886dbf86d02f2
The current monkit reporting for "remote_segments_lost" is not usable for
triggering alerts, as it has reported no data. To allow alerting, two new
metrics "checker_segments_below_min_req" and "repairer_segments_below_min_req"
will increment by zero on each segment unless it is below the minimum
required piece count. The two metrics report what is found by the checker
and the repairer respectively.
Change-Id: I98a68bb189eaf68a833d25cf5db9e68df535b9d7
Make metainfo.RSConfig a valid pflag config value. This allows us to
configure the RSConfig as a string like k/m/o/n-shareSize, which makes
having multiple supported RS schemes easier in the future.
RS-related config values that are no longer needed have been removed
(MinTotalThreshold, MaxTotalThreshold, MaxBufferMem, Verify).
Change-Id: I0178ae467dcf4375c504e7202f31443d627c15e1
A change was made to use a metabase.SegmentKey (a byte slice alias)
as the last seen item to iterate through the irreparable DB in a
for loop. However, this SegmentKey was not initialized, thus it was
nil. This caused the DB query to return nothing, and healthy segments
could not be cleaned out of the irreparable DB.
Change-Id: Idb30d6fef6113a30a27158d548f62c7443e65a81
With the new overlay.AuditOutcome type for offline audits, the
IsUp field is redundant. If AuditOutcome != AuditOffline, then
the node is online.
In addition to removing the field itself, other changes needed
to be made regarding the relationship between 'uptime' and 'audits'.
Previously, uptime and audit outcome were completely separated. For
example, it was possible to update a node's stats to give it a
successful/failed/unknown audit while simultaneously indicating that
the node was offline by setting IsUp to false. This is no longer possible
under this changeset. Some test which did this have been changed slightly
in order to pass.
Also add new benchmarks for UpdateStats and BatchUpdateStats with different
audit outcomes.
Change-Id: I998892d615850b1f138dc62f9b050f720ea0926b
As part of the Metainfo Refactoring, we need to make the Metainfo Loop
working with both the current PointerDB and the new Metabase. Thus, the
Metainfo Loop should pass to the Observer interface more specific Object
and Segment types instead of pb.Pointer.
After this change, there are still a couple of use cases that require
access to the pb.Pointer (hence we have it as a field in the
metainfo.Segment type):
1. Expired Deletion Service
2. Repair Service
It would require additional refactoring in these two services before we
are able to clean this.
Change-Id: Ib3eb6b7507ed89d5ba745ffbb6b37524ef10ed9f
Repair workers prioritize the most unhealthy segments. This has the consequence that when we
finally begin to reach the end of the queue, a good portion of the remaining segments are
healthy again as their nodes have come back online. This makes it appear that there are more
injured segments than there actually are.
solution:
Any time the checker observes an injured segment it inserts it into the repair queue or
updates it if it already exists. Therefore, we can determine which segments are no longer
injured if they were not inserted or updated by the last checker iteration. To do this we
add a new column to the injured segments table, updated_at, which is set to the current time
when a segment is inserted or updated. At the end of the checker iteration, we can delete any
items where updated_at < checker start.
Change-Id: I76a98487a4a845fab2fbc677638a732a95057a94
Another change which is a part of refactoring to replace path parameter
(string/[]byte) with key paramter (metabase.SegmentKey)
Change-Id: I617878442442e5d59bbe5c995f913c3c93c16928
* The audit worker wants to get items from the queue and process them.
* The audit chore wants to create new queues and swap them in when the
old queue has been processed.
This change adds a "Queues" struct which handles the concurrency
issues around the worker fetching a queue and the chore swapping a new
queue in. It simplifies the logic of the "Queue" struct to its bare
bones, so that it behaves like a normal queue with no need to understand
the details of swapping and worker/chore interactions.
Change-Id: Ic3689ede97a528e7590e98338cedddfa51794e1b
This change removes the overlay function FindStorageNodesForRepair,
which skips using the node selection cache and hits the database
directly. Otherwise, it is functionally identical to
FindStorageNodesForUpload, which checks the node selection cache first.
When selecting nodes for PUT_REPAIRs, we now call
FindStorageNodesForUpload instead of FindStorageNodesForRepair to reduce
database load.
Change-Id: If34e109695b2ed2b8fb6759115bf769a3459684e
* Do not swap the active audit queue with the pending audit queue until
the active audit queue is empty.
* Do not begin creating a new pending audit queue until the existing
pending audit queue has been swapped to the active queue.
Change-Id: I81db5bfa01458edb8cdbe71f5baeebdcb1b94317
It feels weird having a repairer configuration part of order services.
Let's have a single source of truth for it.
Change-Id: I24f7c897aec80f3293f8af24876cbb6733d85a0b
Inside CreateGetRepairOrderLimits we pass in a list of healthy pieces,
but when we query node info from this list we apply the "reliable" filter
again. We sometimes end up with nodes which at first were healthy, but then
became unhealthy, and thus can be repaired, but we do not update the 'unhealthyPieces'
list with these nodes.
This causes an error, 'piece to add already exists', as we fail to remove these
pieces from the pointer before replacing them with repaired pieces.
Change-Id: I6e2445f342ac117ded30351fa7e5e523c9ec26bd
What:
Use the github.com/jackc/pgx postgresql driver in place of
github.com/lib/pq.
Why:
github.com/lib/pq has some problems with error handling and context
cancellations (i.e. it might even issue queries or DML statements more
than once! see https://github.com/lib/pq/issues/939). The
github.com/jackx/pgx library appears not to have these problems, and
also appears to be better engineered and implemented (in particular, it
doesn't use "exceptions by panic"). It should also give us some
performance improvements in some cases, and even more so if we can use
it directly instead of going through the database/sql layer.
Change-Id: Ia696d220f340a097dee9550a312d37de14ed2044
Adds monkit tracing for ecrepairer.downloadAndVerifyPiece and
ecrepairer.putPiece so we can get more accurate estimates of node
performance during repair.
Change-Id: Ic05025bf3c493bb3d6f5d325d090c5b7c9e5465d
This will speed up the Put step of repair by not waiting to time out for
a handful of slow nodes, at the expense of a slightly less durable
pointer. It will still repair to the optimal threshold, but not every
node that is selected will end up in the pointer.
Change-Id: I02a0658e3fe6fc0383f26af0f50a065b8b11a651
* add monkit stat new_remote_segments_needing_repair, which reports the
number of new unhealthy segments in the repair queue since the previous
checker iteration
Change-Id: I2f10266006fdd6406ece50f4759b91382059dcc3
