in k8s things really do come online very asynchronously,
we need to use implementation that allows this randomness.
To facilitate this move WriteAll() as part of the
websocket layer instead.
Bonus: avoid instances of dnscache usage on k8s
* Remove lock for cached operations.
* Rename "Relax" to `ReturnLastGood`.
* Add `CacheError` to allow caching values even on errors.
* Add NoWait that will return current value with async fetching if within 2xTTL.
* Make benchmark somewhat representative.
```
Before: BenchmarkCache-12 16408370 63.12 ns/op 0 B/op
After: BenchmarkCache-12 428282187 2.789 ns/op 0 B/op
```
* Remove `storageRESTClient.scanning`. Nonsensical - RPC clients will not have any idea about scanning.
* Always fetch remote diskinfo metrics and cache them. Seems most calls are requesting metrics.
* Do async fetching of usage caches.
- bucket metadata does not need to look for legacy things
anymore if b.Created is non-zero
- stagger bucket metadata loads across lots of nodes to
avoid the current thundering herd problem.
- Remove deadlines for RenameData, RenameFile - these
calls should not ever be timed out and should wait
until completion or wait for client timeout. Do not
choose timeouts for applications during the WRITE phase.
- increase R/W buffer size, increase maxMergeMessages to 30
Allow internal types to support a `Recycler` interface, which will allow for sharing of common types across handlers.
This means that all `grid.MSS` (and similar) objects are shared across in a common pool instead of a per-handler pool.
Add internal request reuse of internal types. Add for safe (pointerless) types explicitly.
Only log params for internal types. Doing Sprint(obj) is just a bit too messy.
Each Put, List, Multipart operations heavily rely on making
GetBucketInfo() call to verify if bucket exists or not on
a regular basis. This has a large performance cost when there
are tons of servers involved.
We did optimize this part by vectorizing the bucket calls,
however its not enough, beyond 100 nodes and this becomes
fairly visible in terms of performance.
- Move RenameFile to websockets
- Move ReadAll that is primarily is used
for reading 'format.json' to to websockets
- Optimize DiskInfo calls, and provide a way
to make a NoOp DiskInfo call.
NOTE: This feature is not retro-active; it will not cater to previous transactions
on existing setups.
To enable this feature, please set ` _MINIO_DRIVE_QUORUM=on` environment
variable as part of systemd service or k8s configmap.
Once this has been enabled, you need to also set `list_quorum`.
```
~ mc admin config set alias/ api list_quorum=auto`
```
A new debugging tool is available to check for any missing counters.
Bonus: allow replication to attempt Deletes/Puts when
the remote returns quorum errors of some kind, this is
to ensure that MinIO can rewrite the namespace with the
latest version that exists on the source.
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.
replace io.Discard usage to fix NUMA copy() latencies
On NUMA systems copying from 8K buffer allocated via
io.Discard leads to large latency build-up for every
```
copy(new8kbuf, largebuf)
```
can in-cur upto 1ms worth of latencies on NUMA systems
due to memory sharding across NUMA nodes.
.metacache objects are transient in nature, and are better left to
use page-cache effectively to avoid using more IOPs on the disks.
this allows for incoming calls to be not taxed heavily due to
multiple large batch listings.
health checks were missing for drives replaced since
- HealFormat() would replace the drives without a health check
- disconnected drives when they reconnect via connectEndpoint()
the loop also loses health checks for local disks and merges
these into a single code.
- other than this separate cleanUp, health check variables to avoid
overloading them with similar requirements.
- also ensure that we compete via context selector for disk monitoring
such that the canceled disks don't linger around longer waiting for
the ticker to trigger.
- allow disabling active monitoring.
* Reduce allocations
* Add stringsHasPrefixFold which can compare string prefixes, while ignoring case and not allocating.
* Reuse all msgp.Readers
* Reuse metadata buffers when not reading data.
* Make type safe. Make buffer 4K instead of 8.
* Unslice
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.
xl.meta gets written and never rolled back, however
we definitely need to validate the state that is
persisted on the disk, if there are inconsistencies
- more than write quorum we should return an error
to the client
- if write quorum was achieved however there are
inconsistent xl.meta's we should simply trigger
an MRF on them
Do completely independent multipart uploads.
In distributed mode, a lock was held to merge each multipart
upload as it was added. This lock was highly contested and
retries are expensive (timewise) in distributed mode.
Instead, each part adds its metadata information uniquely.
This eliminates the per object lock required for each to merge.
The metadata is read back and merged by "CompleteMultipartUpload"
without locks when constructing final object.
Co-authored-by: Harshavardhana <harsha@minio.io>
fix: change timedvalue to return previous cached value
caller can interpret the underlying error and decide
accordingly, places where we do not interpret the
errors upon timedValue.Get() - we should simply use
the previously cached value instead of returning "empty".
Bonus: remove some unused code
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.
deleting collection of versions belonging
to same object, we can avoid re-reading
the xl.meta from the disk instead purge
all the requested versions in-memory,
the tradeoff is to allocate a map to de-dup
the versions, allow disks to be read only
once per object.
additionally reduce the data transfer between
nodes by shortening msgp data values.
Harshavardhana
Aditya Manthramurthy
Klaus Post
Anis Eleuch
jiuker
Praveen raj Mani