epoll contention on TCP causes latency build-up when
we have high volume ingress. This PR is an attempt to
relieve this pressure.
upstream issue https://github.com/golang/go/issues/65064
It seems to be a deeper problem; haven't yet tried the fix
provide in this issue, but however this change without
changing the compiler helps.
Of course, this is a workaround for now, hoping for a
more comprehensive fix from Go runtime.
This PR adds a WebSocket grid feature that allows servers to communicate via
a single two-way connection.
There are two request types:
* Single requests, which are `[]byte => ([]byte, error)`. This is for efficient small
roundtrips with small payloads.
* Streaming requests which are `[]byte, chan []byte => chan []byte (and error)`,
which allows for different combinations of full two-way streams with an initial payload.
Only a single stream is created between two machines - and there is, as such, no
server/client relation since both sides can initiate and handle requests. Which server
initiates the request is decided deterministically on the server names.
Requests are made through a mux client and server, which handles message
passing, congestion, cancelation, timeouts, etc.
If a connection is lost, all requests are canceled, and the calling server will try
to reconnect. Registered handlers can operate directly on byte
slices or use a higher-level generics abstraction.
There is no versioning of handlers/clients, and incompatible changes should
be handled by adding new handlers.
The request path can be changed to a new one for any protocol changes.
First, all servers create a "Manager." The manager must know its address
as well as all remote addresses. This will manage all connections.
To get a connection to any remote, ask the manager to provide it given
the remote address using.
```
func (m *Manager) Connection(host string) *Connection
```
All serverside handlers must also be registered on the manager. This will
make sure that all incoming requests are served. The number of in-flight
requests and responses must also be given for streaming requests.
The "Connection" returned manages the mux-clients. Requests issued
to the connection will be sent to the remote.
* `func (c *Connection) Request(ctx context.Context, h HandlerID, req []byte) ([]byte, error)`
performs a single request and returns the result. Any deadline provided on the request is
forwarded to the server, and canceling the context will make the function return at once.
* `func (c *Connection) NewStream(ctx context.Context, h HandlerID, payload []byte) (st *Stream, err error)`
will initiate a remote call and send the initial payload.
```Go
// A Stream is a two-way stream.
// All responses *must* be read by the caller.
// If the call is canceled through the context,
//The appropriate error will be returned.
type Stream struct {
// Responses from the remote server.
// Channel will be closed after an error or when the remote closes.
// All responses *must* be read by the caller until either an error is returned or the channel is closed.
// Canceling the context will cause the context cancellation error to be returned.
Responses <-chan Response
// Requests sent to the server.
// If the handler is defined with 0 incoming capacity this will be nil.
// Channel *must* be closed to signal the end of the stream.
// If the request context is canceled, the stream will no longer process requests.
Requests chan<- []byte
}
type Response struct {
Msg []byte
Err error
}
```
There are generic versions of the server/client handlers that allow the use of type
safe implementations for data types that support msgpack marshal/unmarshal.
PUT calls cannot afford to have large latency build-ups due
to contentious usage.json, or worse letting them fail with
some unexpected error, this can happen when this file is
concurrently being updated via scanner or it is being
healed during a disk replacement heal.
However, these are fairly quick in theory, stressed clusters
can quickly show visible latency this can add up leading to
invalid errors returned during PUT.
It is perhaps okay for us to relax this error return requirement
instead, make sure that we log that we are proceeding to take in
the requests while the quota is using an older value for the quota
enforcement. These things will reconcile themselves eventually,
via scanner making sure to overwrite the usage.json.
Bonus: make sure that storage-rest-client sets ExpectTimeouts to
be 'true', such that DiskInfo() call with contextTimeout does
not prematurely disconnect the servers leading to a longer
healthCheck, back-off routine. This can easily pile up while also
causing active callers to disconnect, leading to quorum loss.
DiskInfo is actively used in the PUT, Multipart call path for
upgrading parity when disks are down, it in-turn shouldn't cause
more disks to go down.
Since JWT tokens remain valid for up to 15 minutes, we
don't have to regenerate tokens for every call.
Cache tokens for matching access+secret+audience
for up to 15 seconds.
```
BenchmarkAuthenticateNode/uncached-32 270567 4179 ns/op 2961 B/op 33 allocs/op
BenchmarkAuthenticateNode/cached-32 7684824 157.5 ns/op 48 B/op 1 allocs/op
```
Reduces internode call allocations a great deal.
it would seem like using `bufio.Scan()` is very
slow for heavy concurrent I/O, ie. when r.Body
is slow , instead use a proper
binary exchange format, to marshal and unmarshal
the LockArgs datastructure in a cleaner way.
this PR increases performance of the locking
sub-system for tiny repeated read lock requests
on same object.
```
BenchmarkLockArgs
BenchmarkLockArgs-4 6417609 185.7 ns/op 56 B/op 2 allocs/op
BenchmarkLockArgsOld
BenchmarkLockArgsOld-4 1187368 1015 ns/op 4096 B/op 1 allocs/op
```
This is to ensure that there are no projects
that try to import `minio/minio/pkg` into
their own repo. Any such common packages should
go to `https://github.com/minio/pkg`
This refactor is done for few reasons below
- to avoid deadlocks in scenarios when number
of nodes are smaller < actual erasure stripe
count where in N participating local lockers
can lead to deadlocks across systems.
- avoids expiry routines to run 1000 of separate
network operations and routes per disk where
as each of them are still accessing one single
local entity.
- it is ideal to have since globalLockServer
per instance.
- In a 32node deployment however, each server
group is still concentrated towards the
same set of lockers that partipicate during
the write/read phase, unlike previous minio/dsync
implementation - this potentially avoids send
32 requests instead we will still send at max
requests of unique nodes participating in a
write/read phase.
- reduces overall chattiness on smaller setups.
fixes a regression introduced in #10859, due
to the error returned by rest.Client being typed
i.e *rest.NetworkError - IsNetworkHostDown function
didn't work as expected to detect network issues.
This in-turn aggravated the situations when nodes
are disconnected leading to performance loss.
This will make the health check clients 'silent'.
Use `IsNetworkOrHostDown` determine if network is ok so it mimics the functionality in the actual client.
lockers currently might leave stale lockers,
in unknown ways waiting for downed lockers.
locker check interval is high enough to safely
cleanup stale locks.
- select lockers which are non-local and online to have
affinity towards remote servers for lock contention
- optimize lock retry interval to avoid sending too many
messages during lock contention, reduces average CPU
usage as well
- if bucket is not set, when deleteObject fails make sure
setPutObjHeaders() honors lifecycle only if bucket name
is set.
- fix top locks to list out always the oldest lockers always,
avoid getting bogged down into map's unordered nature.
- Add owner information for expiry, locking, unlocking a resource
- TopLocks returns now locks in quorum by default, provides
a way to capture stale locks as well with `?stale=true`
- Simplify the quorum handling for locks to avoid from storage
class, because there were challenges to make it consistent
across all situations.
- And other tiny simplifications to reset locks.
Add context to all (non-trivial) calls to the storage layer.
Contexts are propagated through the REST client.
- `context.TODO()` is left in place for the places where it needs to be added to the caller.
- `endWalkCh` could probably be removed from the walkers, but no changes so far.
The "dangerous" part is that now a caller disconnecting *will* propagate down, so a
"delete" operation will now be interrupted. In some cases we might want to disconnect
this functionality so the operation completes if it has started, leaving the system in a cleaner state.
Context timeout might race on each other when timeouts are lower
i.e when two lock attempts happened very quickly on the same resource
and the servers were yet trying to establish quorum.
This situation can lead to locks held which wouldn't be unlocked
and subsequent lock attempts would fail.
This would require a complete server restart. A potential of this
issue happening is when server is booting up and we are trying
to hold a 'transaction.lock' in quick bursts of timeout.
Without instantiating a new rest client we can
have a recursive error which can lead to
healthcheck returning always offline, this can
prematurely take the servers offline.
- Add changes to ensure remote disks are not
incorrectly taken online if their order has
changed or are incorrect disks.
- Bring changes to peer to detect disconnection
with separate Health handler, to avoid a
rather expensive call GetLocakDiskIDs()
- Follow up on the same changes for Lockers
as well
At a customer setup with lots of concurrent calls
it can be observed that in newRetryTimer there
were lots of tiny alloations which are not
relinquished upon retries, in this codepath
we were only interested in re-using the timer
and use it wisely for each locker.
```
(pprof) top
Showing nodes accounting for 8.68TB, 97.02% of 8.95TB total
Dropped 1198 nodes (cum <= 0.04TB)
Showing top 10 nodes out of 79
flat flat% sum% cum cum%
5.95TB 66.50% 66.50% 5.95TB 66.50% time.NewTimer
1.16TB 13.02% 79.51% 1.16TB 13.02% github.com/ncw/directio.AlignedBlock
0.67TB 7.53% 87.04% 0.70TB 7.78% github.com/minio/minio/cmd.xlObjects.putObject
0.21TB 2.36% 89.40% 0.21TB 2.36% github.com/minio/minio/cmd.(*posix).Walk
0.19TB 2.08% 91.49% 0.27TB 2.99% os.statNolog
0.14TB 1.59% 93.08% 0.14TB 1.60% os.(*File).readdirnames
0.10TB 1.09% 94.17% 0.11TB 1.25% github.com/minio/minio/cmd.readDirN
0.10TB 1.07% 95.23% 0.10TB 1.07% syscall.ByteSliceFromString
0.09TB 1.03% 96.27% 0.09TB 1.03% strings.(*Builder).grow
0.07TB 0.75% 97.02% 0.07TB 0.75% path.(*lazybuf).append
```
Change distributed locking to allow taking bulk locks
across objects, reduces usually 1000 calls to 1.
Also allows for situations where multiple clients sends
delete requests to objects with following names
```
{1,2,3,4,5}
```
```
{5,4,3,2,1}
```
will block and ensure that we do not fail the request
on each other.
Use reference format to initialize lockers
during startup, also handle `nil` for NetLocker
in dsync and remove *errorLocker* implementation
Add further tuning parameters such as
- DialTimeout is now 15 seconds from 30 seconds
- KeepAliveTimeout is not 20 seconds, 5 seconds
more than default 15 seconds
- ResponseHeaderTimeout to 10 seconds
- ExpectContinueTimeout is reduced to 3 seconds
- DualStack is enabled by default remove setting
it to `true`
- Reduce IdleConnTimeout to 30 seconds from
1 minute to avoid idleConn build up
Fixes#8773
This PR implements locking from a global entity into
a more localized set level entity, allowing for locks
to be held only on the resources which are writing
to a collection of disks rather than a global level.
In this process this PR also removes the top-level
limit of 32 nodes to an unlimited number of nodes. This
is a precursor change before bring in bucket expansion.
In distributed mode, use REST API to acquire and manage locks instead
of RPC.
RPC has been completely removed from MinIO source.
Since we are moving from RPC to REST, we cannot use rolling upgrades as the
nodes that have not yet been upgraded cannot talk to the ones that have
been upgraded.
We expect all minio processes on all nodes to be stopped and then the
upgrade process to be completed.
Also force http1.1 for inter-node communication
Harshavardhana
Klaus Post
Anis Elleuch
Krishna Srinivas
kannappanr