ClassifySegmentPieces uses custom set implementation instead map.
Side note, for custom set implementation I also checked int8 bit set but
it didn't give better performance so I used simpler implementation.
Benchmark results (compared against part 2 optimization change):
name old time/op new time/op delta
RemoteSegment/healthy_segment-8 21.7µs ± 8% 15.4µs ±16% -29.38% (p=0.008 n=5+5)
name old alloc/op new alloc/op delta
RemoteSegment/healthy_segment-8 7.41kB ± 0% 1.87kB ± 0% -74.83% (p=0.000 n=5+4)
name old allocs/op new allocs/op delta
RemoteSegment/healthy_segment-8 150 ± 0% 130 ± 0% -13.33% (p=0.008 n=5+5)
Change-Id: I21feca9ec6ac0a2558ac5ce8894451c54f69e52d
It would appear that we have been making concurrent accesses to
statsCollector for a long, long time (we expect there to be multiple
calls to `Repair()` at the same time on the same instance of
`SegmentRepairer`, up to `config.MaxRepair`, and before this change
there was no sort of synchronization guarding accesses to the
`statsCollector.stats` map.
Refs: https://github.com/storj/storj/issues/6402
Change-Id: I5bcdd13c88913a8d66f6dd906c9037c588960cc9
This patch finishes the placement aware repair.
We already introduced the parameters to select only the jobs for specific placements, the remaining part is just to configure the exclude/include rules. + a full e2e unit test.
Change-Id: I223ba84e8ab7481a53e5a444596c7a5ae51573c5
Currently, graceful exit is a complicated subsystem that keeps a queue
of all pieces expected to be on a node, and asks the node to transfer
those pieces to other nodes one by one. The complexity of the system
has, unfortunately, led to numerous bugs and unexpected behaviors.
We have decided to remove this entire subsystem and restructure graceful
exit as follows:
* Nodes will signal their intent to exit gracefully
* The satellite will not send any new pieces to gracefully exiting nodes
* Pieces on gracefully exiting nodes will be considered by the repair
subsystem as "retrievable but unhealthy". They will be repaired off of
the exiting node as needed.
* After one month (with an appropriately high online score), the node
will be considered exited, and held amounts for the node will be
released. The repair worker will continue to fetch pieces from the
node as long as the node stays online.
* If, at the end of the month, a node's online score is below a certain
threshold, its graceful exit will fail.
Refs: https://github.com/storj/storj/issues/6042
Change-Id: I52d4e07a4198e9cb2adf5e6cee2cb64d6f9f426b
The repair checker and repair worker both need to determine which pieces
are healthy, which are retrievable, and which should be replaced, but
they have been doing it in different ways in different code, which has
been the cause of bugs. The same term could have very similar but subtly
different meanings between the two, causing much confusion.
With this change, the piece- and node-classification logic is
consolidated into one place within the satellite/repair package, so that
both subsystems can use it. This ought to make decision-making code more
concise and more readable.
The consolidated classification logic has been expanded to create more
sets, so that the decision-making code does not need to do as much
precalculation. It should now be clearer in comments and code that a
piece can belong to multiple sets arbitrarily (except where the
definition of the sets makes this logically impossible), and what the
precise meaning of each set is. These sets include Missing, Suspended,
Clumped, OutOfPlacement, InExcludedCountry, ForcingRepair,
UnhealthyRetrievable, Unhealthy, Retrievable, and Healthy.
Some other side effects of this change:
* CreatePutRepairOrderLimits no longer needs to special-case excluded
countries; it can just create as many order limits as requested (by
way of len(newNodes)).
* The repair checker will now queue a segment for repair when there are
any pieces out of placement. The code calls this "forcing a repair".
* The checker.ReliabilityCache is now accessed by way of a GetNodes()
function similar to the one on the overlay. The classification methods
like MissingPieces(), OutOfPlacementPieces(), and
PiecesNodesLastNetsInOrder() are removed in favor of the
classification logic in satellite/repair/classification.go. This
means the reliability cache no longer needs access to the placement
rules or excluded countries list.
Change-Id: I105109fb94ee126952f07d747c6e11131164fadb
When we do `satellite run api --placement '...'`, the placement rules are not parsed well.
The problem is based on `viper.AllSettings()`, and the main logic is sg. like this (from a new unit test):
```
r := ConfigurablePlacementRule{}
err := r.Set(p)
require.NoError(t, err)
serialized := r.String()
r2 := ConfigurablePlacementRule{}
err = r2.Set(serialized)
require.NoError(t, err)
require.Equal(t, p, r2.String())
```
All settings evaluates the placement rules in `ConfigurablePlacementRules` and stores the string representation.
The problem is that we don't have proper `String()` implementation (it prints out the structs instead of the original definition.
There are two main solutions for this problem:
1. We can fix the `String()`. When we parse a placement rule, the `String()` method should print out the original definition
2. We can switch to use pure string as configuration parameter, and parse the rules only when required.
I feel that 1 is error prone, we can do it (and in this patch I added a lot of `String()` implementations, but it's hard to be sure that our `String()` logic is inline with the parsing logic.
Therefore I decided to make the configuration value of the placements a string (or a wrapper around string).
That's the main reason why this patch seems to be big, as I updated all the usages.
But the main part is in beginning of the `placement.go` (configuration parsing is not a pflag.Value implementation any more, but a separated step).
And `filter.go`, (a few more String implementation for filters.
https://github.com/storj/storj/issues/6248
Change-Id: I47c762d3514342b76a2e85683b1c891502a0756a
as GetParticipatingNodes and GetNodes, respectively.
We now want these functions to include offline and suspended nodes as
well, so that we can force immediate repair when pieces are out of
placement or in excluded countries. With that change, the old names no
longer made sense.
Change-Id: Icbcbad43dbde0ca8cbc80a4d17a896bb89b078b7
As I learned, the `Include` supposed to communicate that some internal change also "included" to the filters during the check -> filters might be stateful.
But it's not the case any more after 552242387, where we removed the only one stateful filter.
Change-Id: I7c36ddadb2defbfa3b6b67bcc115e4427ba9e083
This patch is a oneliner: rangedloop checker should check the subnets only if it's not turned off with placement annotation.
(see in satellite/repair/checker/observer.go).
But I didn't find any unit test to cover that part, so I had to write one, and I prefered to write it as a unit test not an integration test, which requires a mock repair queue (observer_unit_test.go mock.go).
Because it's small change, I also included a small change: creating a elper method to check if AutoExcludeSubnet annotation is defined
Change-Id: I2666b937074ab57f603b356408ef108cd55bd6fd
When we check the availability of the pieces, we do:
```
result.NumUnhealthyRetrievable = len(result.ClumpedPiecesSet) + len(result.OutOfPlacementPiecesSet)
// + some magic if there are overlaps between them
numHealthy := len(pieces) - len(piecesCheck.MissingPiecesSet) - piecesCheck.NumUnhealthyRetrievable
```
This works only if OutOfPlacementPieceSet doesn't contain the offline nodes (which are already included in MissingPieceSet).
But `result.OutOfPlacementPieces.Set` should include all the nodes (even offline), as in case of lucky conditions, we are able to remove those pieces from DB.
The solution is to remove all offline nodes from `NumUnhealthyRetrievable`.
Change-Id: I90baa0396352dd040e1e1516314b3271f8712034
This patch fixes the node tag based placement of rangedloop/repairchecker + repair process.
The main change is just adding the node tags for Reliable and KnownReliabel database calls + adding new tests to prove, it works.
https://github.com/storj/storj/issues/6126
Change-Id: I245d654a18c1d61b2c72df49afa0718d0de76da1
There are cases when we would like to override the default placement=0 rule.
For example when we would like to exclude tagged nodes from the selection (by default).
Therefore we couldn't use a shortcut any more, we should always check the placement rules, even if we use placement=0.
TODO: we need to update common, and rename `EveryCountry` to `DefaultPlacement`, just to avoid confusion.
https://github.com/storj/storj/issues/6126
Change-Id: Iba6c655bd623e04351ea7ff91fd741785dc193e4
This feature flag was disabled by default to test it slowly. Its enabled
for some time on one production satellite and test satellites without
any issue. We can enable it by default in code.
Change-Id: If9c36895bbbea12bd4aefa30cb4df912e1729e4c
Sometimes DownloadSelectionCache doesn't keep up with all node
placement changes we are doing during this test.
Change-Id: Idbda6511e3324b560cee3be85f980bf8d5b9b7ef
Currently, we have issue were while counting unhealthy pieces we are
counting twice piece which is in excluded country and is outside segment
placement. This can cause unnecessary repair.
This change is also doing another step to move RepairExcludedCountryCodes
from overlay config into repair package.
Change-Id: I3692f6e0ddb9982af925db42be23d644aec1963f
placement.AllowedCountry is the old way to specify placement, with the new approach we can use a more generic (dynamic method), which can check full node information instead of just the country code.
The 90% of this patch is just search and replace:
* we need to use NodeFilters instead of placement.AllowedCountry
* which means, we need an initialized PlacementRules available everywhere
* which means we need to configure the placement rules
The remaining 10% is the placement.go, where we introduced a new type of configuration (lightweight expression language) to define any kind of placement without code change.
Change-Id: Ie644b0b1840871b0e6bbcf80c6b50a947503d7df
All the files in uploadselection are (in fact) related to generic node selection, and used not only for upload,
but for download, repair, etc...
Change-Id: Ie4098318a6f8f0bbf672d432761e87047d3762ab
We use two different Node types in `overlay` and `uploadnodeselection` and converting back and forth.
Using the same object would allow us to use a unified node selection interface everywhere.
Change-Id: Ie71e29d60184ee0e5b4547eb54325f09c418f73c
At the moment segment repairer is skipping offline nodes in checks like
clumped pieces and off placement pieces. This change is fixing this
problem using new version of KnownReliable method. New method is
returning both online and offline nodes. Provided data can be used to
find clumped and off placement pieces.
We are not using DownloadSelectionCache anymore with segment repairer.
https://github.com/storj/storj/issues/5998
Change-Id: I236a1926e21f13df4cdedc91130352d37ff97e18
When pieces fail an audit (hard fail, meaning the node acknowledged it
did not have the piece or the piece was corrupted), we will now remove
those pieces from the segment.
Previously, we did not do this, and some node operators were seeing the
same missing piece audited over and over again and losing reputation
every time.
This change will include both verification and reverification audits. It
will also apply to pieces found to be bad during repair, if
repair-to-reputation reporting is enabled.
Change-Id: I0ca7af7e3fecdc0aebbd34fee4be3a0eab53f4f7
Additional test case to cover situation where we are trying to
repair segment with specific placement set. We need to be sure
that segment won't be repaired into nodes that are outside
segment placement, even if that means that repair will fail.
Change-Id: I99d238aa9d9b9606eaf89cd1cf587a2585faee91
This change makes dial timeout configurable and change it also from
defatul 20s to 5s. Main motivation is that during repair we often loose
lots of time to dial which eventually will fail. New timeout should be
still enough to dial but we will move forward quicker to next node if
that one will fail.
Timeout is also applied directly as context timeout in case we will
use noise of tcp fast open one day.
Change-Id: I021bf459af49b11241e314fa1a7887c81d5214ea
We missed to set placement as a part of selection request. It can case
uploading repaired data out of specified placement.
I will provide test as a separate change.
Change-Id: I4efe67f2d5f545a1d70e831e5d297f0977a4eed1
We were reusing a slice to save on allocations, but it turns out the
function using it was being called in multiple goroutines at the same
time.
This is definitely a problem with repairer/segments.go. I'm not 100%
sure if it also is a problem with checker/observer.go, but I'm making
the change there as well to be on the safe side for now.
Repair workers only ran with this bug on testing satellites, and it
looks like the worst that could have happened was that we repaired
pieces off of well-behaved, non-clumped, in-placement nodes by mistake.
Change-Id: I33c112b05941b63d066caab6a34a543840c6b85d
Segment repairer should take into account segment 'placement' field
and remove or repair pieces from nodes that are outside this placement.
In case when after considering pieces out of placement we are still above
repair threshold we are only updating segment pieces to remove
problematic pieces. Otherwise we are doing regular repair.
https://github.com/storj/storj/issues/5896
Change-Id: I72b652aff2e6b20be3ac6dbfb1d32c2840ce3d59
It seems that the "what pieces are clumped" code does not work right, so
this logic is causing repair overload or other repair failures.
Hide it behind a flag while we figure out what is going on, so that
repair can still work in the meantime.
Change-Id: If83ef7895cba870353a67ab13573193d92fff80b
We avoid putting more than one piece of a segment on the same /24
network (or /64 for ipv6). However, it is possible for multiple pieces
of the same segment to move to the same network over time. Nodes can
change addresses, or segments could be uploaded with dev settings, etc.
We will call such pieces "clumped", as they are clumped into the same
net, and are much more likely to be lost or preserved together.
This change teaches the repair checker to recognize segments which have
clumped pieces, and put them in the repair queue. It also teaches the
repair worker to repair such segments (treating clumped pieces as
"retrievable but unhealthy"; i.e., they will be replaced on new nodes if
possible).
Refs: https://github.com/storj/storj/issues/5391
Change-Id: Iaa9e339fee8f80f4ad39895438e9f18606338908
The following tests should be made less flaky by this change:
- TestFailedDataRepair
- TestOfflineNodeDataRepair
- TestUnknownErrorDataRepair
- TestMissingPieceDataRepair_Succeed
- TestMissingPieceDataRepair
- TestCorruptDataRepair_Succeed
- TestCorruptDataRepair_Failed
This follows on to a change in commit 6bb64796. Nearly all tests in the
repair suite used to rely on events happening in a certain order. After
some of our performance work, those things no longer happen in that
expected order every time. This caused much flakiness.
The fix in 6bb64796 was sufficient for the tests operating directly on
an `*ECRepairer` instance, but not for the tests that make use of the
repairer by way of the repair queue and the repair worker. These tests
needed a different way to indicate the number of expected failures. This
change provides that different way.
Refs: https://github.com/storj/storj/issues/5736
Refs: https://github.com/storj/storj/issues/5718
Refs: https://github.com/storj/storj/issues/5715
Refs: https://github.com/storj/storj/issues/5609
Change-Id: Iddcf5be3a3ace7ad35fddb513ab53dd3f2f0eb0e
It was possible to get into a situation where successfulPieces =
es.RequiredCount(), errorCount < minFailures, and inProgress == 0 (when
the succeeding gets all completed before the failures), whereupon the
last goroutine in the limiter would sit and wait forever for another
goroutine to finish.
This change corrects the handling of that situation.
As an aside, this is really pretty confusing code and we should think
about redoing the whole function.
Change-Id: Ifa3d3ad92bc755e563fd06b2aa01ef6147075a69
This code is essentially replacement for eestream.CalcPieceSize. To call
eestream.CalcPieceSize we need eestream.RedundancyStrategy which is not
trivial to get as it requires infectious.FEC. For example infectious.FEC
creation is visible on GE loop observer CPU profile because we were
doing this for each segment in DB.
New method was added to storj.Redundancy and here we are just wiring it
with metabase Segment.
BenchmarkSegmentPieceSize
BenchmarkSegmentPieceSize/eestream.CalcPieceSize
BenchmarkSegmentPieceSize/eestream.CalcPieceSize-8 5822 189189 ns/op 9776 B/op 8 allocs/op
BenchmarkSegmentPieceSize/segment.PieceSize
BenchmarkSegmentPieceSize/segment.PieceSize-8 94721329 11.49 ns/op 0 B/op 0 allocs/op
Change-Id: I5a8b4237aedd1424c54ed0af448061a236b00295
Several tests using `(*ECRepairer).Get()` have begun to exhibit flaky
results. The tests are expecting to see failures in certain cases, but
the failures are not present. It appears that the cause of this is that,
sometimes, the fastest good nodes are able to satisfy the repairer
(providing RequiredCount pieces) before the repairer is able to identify
the problem scenario we have laid out.
In this commit, we add an argument to `(*ECRepairer).Get()` which
specifies how many failure results are expected. In normal/production
conditions, this parameter will be 0, meaning Get need not wait for
any errors and should only report those that arrived while waiting for
RequiredCount pieces (the existing behavior). But in these tests, we can
request that Get() wait for enough results to see the errors we are
expecting.
Refs: https://github.com/storj/storj/issues/5593
Change-Id: I2920edb6b5a344491786aab794d1be6372c07cf8
implemented observer and partial, created new structures to keep mon
metrics remain in same way as in segment loop
Change-Id: I209c126096c84b94d4717332e56238266f6cd004
The Reporter is responsible for processing results from auditing
operations, logging the results, disqualifying nodes that reached
the maximum reverification count, and passing the results on to
the reputation system.
In this commit, we extend the Reporter so that it knows how to process
the results of piecewise reverification audits.
We also change most reporter-related tests so that reverifications
happen as piecewise reverification audits, exercising the new code.
Note that piecewise reverification audits are not yet being done outside
of tests. In a later commit, we will switch from doing segmentwise
reverifications to piecewise reverifications, as part of the
audit-scaling effort.
Refs: https://github.com/storj/storj/issues/5230
Change-Id: I9438164ce1ea4d9a1790d18d0e1046a8eb04d8e9
Reputation updates during repair currently consumes a lot of database
resources. Sometimes increasing the rate of repair is more important
than auditing a node based on whether they have or don't have the
correct piece during repair. This is the job of the audit service.
This commit is to implement an intermediate solution from this issue: https://github.com/storj/storj/issues/5089
This commit does not address the more in-depth fix discussed here: https://github.com/storj/storj/issues/4939
Change-Id: I4163b18d78a96fadf5265789fd73c8aa8def0e9f
We have an alert on `repair_too_many_nodes_failed` which fires too
frequently. Every time so far, it has been because of a network blip of
some nature on the satellite side.
Satellite operators are expected to have other means in place for
alerting on network problems and fixing them, so it's not necessary for
the repair framework to act in that way.
Instead, in this change, we change the way that
`repair_too_many_nodes_failed` works. When a repair fails, we collect
piece fetch errors by type and determine from them whether it looks like
we are having network problems (most errors are connection failures,
possibly also some successful connections which subsequently time out)
or whether something else has happened.
We will now only emit `repair_too_many_nodes_failed` when the outcome
does not look like a network failure. In the network failure case, we
will instead emit `repair_suspected_network_problem`.
Refs: https://github.com/storj/storj/issues/4669
Change-Id: I49df98da5df9c606b95ad08a2bdfec8092fba926
This structure is entirely unused within the audit module, and is only
used by repair code. Accordingly, this change moves the structure from
audit code to repair code.
Also, we take the opportunity here to rename the structure to something
less generic.
Refs: https://github.com/storj/storj/issues/4669
Change-Id: If85b37e08620cda1fde2afe98206293e02b5c36e
We retry a GET_REPAIR operation in one case, and one case only (as far
as I can determine): when we are trying to connect to a node using its
last known working IP and port combination rather than its supplied
hostname, and we think the operation failed the first time because of a
Dial failure.
However, logs collected from storage node operators along with logs
collected from satellites are strongly indicating that we are retrying
GET_REPAIR operations in some cases even when we succeeded in connecting
to the node the first time. This results in the node complaining loudly
about being given a duplicate order limit (as it should), whereupon the
satellite counts that as an unknown error and potentially penalizes the
node.
See discussion at
https://forum.storj.io/t/get-repair-error-used-serial-already-exists-in-store/17922/36
.
Investigation into this problem has revealed that
`!piecestore.CloseError.Has(err)` may not be the best way of determining
whether a problem occurred during Dial. In fact, it is probably
downright Wrong. Handling of errors on a stream is somewhat complicated,
but it would appear that there are several paths by which an RPC error
originating on the remote side might show up during the Close() call,
and would thus be labeled as a "CloseError".
This change creates a new error class, repairer.ErrDialFailed, with
which we will now wrap errors that _really definitely_ occurred during
a Dial call. We will use this class to determine whether or not to retry
a GET_REPAIR operation. The error will still also be wrapped with
whatever wrapper classes it used to be wrapped with, so the potential
for breakage here should be minimal.
Refs: https://github.com/storj/storj/issues/4687
Change-Id: Ifdd3deadc8258f34cf3fbc42aff393fa545794eb
To save load on DNS servers, the repair code first tries to dial the
last known good ip and port for a node, and then falls back to a DNS
lookup only if we fail to connect to the last known good ip and port.
However, it looks like we are seeing errors during the client stream
Close() call (probably due to quic-go code), and those are classified
the same as errors encountered during Dial. The repairer code sees this
error, assumes that we failed to contact the node, and retries- but
since we did actually succeed in connecting the first time around, this
results in submitting the same order limit (with the same serial number)
to the storage node, which (rightfully) rejects it.
So together with change I055c186d5fd4e79560f67763175bc3130b9bc7d2 in
storj/uplink, this should avoid the double submission and avoid dinging
nodes' suspension scores unfairly.
See https://github.com/storj/storj/issues/4687.
Also, moving the testsuite directory check up above check-monkit in the
Jenkins Lint task, so that a non-tidy testsuite/go.mod can be recognized
and handled before everything breaks weirdly and seemingly randomly
later on.
Change-Id: Icb2b05aaff921d0af6aba10e450ac7e0a7bb2655
