The path is marked dirty automatically when healObject() is called, which is
wrong. HealObject() is called during self-healing and this will lead to
an increase in the false positive result of the bloom filter.
Also move NSUpdated() from renameData() and call it directly in
CompleteMultipart and PutObject, this is not a functional change but
it will make it less prone to errors in the future.
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>
Uploading a part object can leave an inconsistent state inside
.minio.sys/multipart where data are uploaded but xl.meta is not
committed yet.
Do not list upload-ids that have this state in the multipart listing.
We need to make sure if we cannot read bucket metadata
for some reason, and bucket metadata is not missing and
returning corrupted information we should panic such
handlers to disallow I/O to protect the overall state
on the system.
In-case of such corruption we have a mechanism now
to force recreate the metadata on the bucket, using
`x-minio-force-create` header with `PUT /bucket` API
call.
Additionally fix the versioning config updated state
to be set properly for the site replication healing
to trigger correctly.
do not modify opts.UserDefined after object-handler
has set all the necessary values, any mutation needed
should be done on a copy of this value not directly.
As there are other pieces of code that access opts.UserDefined
concurrently this becomes challenging.
fixes#14856
When reading input for PutObject or PutObjectPart add a readahead buffer for big inputs.
This will make network reads+hashing separate run async with erasure coding and writes. This will reduce overall latency in distributed setups where the input is from upstream and writes go to other servers.
We will read at 2 buffers ahead, meaning one will always be ready/waiting and one is currently being read from.
This improves PutObject and PutObjectParts for these cases.
This PR simplifies few things
- Multipart parts are renamed, upon failure are unrenamed() keep this
multipart specific behavior it is needed and works fine.
- AbortMultipart should blindly delete once lock is acquired instead
of re-reading metadata and calculating quorum, abort is a delete()
operation and client has no business looking for errors on this.
- Skip Access() calls to folders that are operating on
`.minio.sys/multipart` folder as well.
dataDir loosely based on maxima is incorrect and does not
work in all situations such as disks in the following order
- xl.json migration to xl.meta there may be partial xl.json's
leftover if some disks are not yet connected when the disk
is yet to come up, since xl.json mtime and xl.meta is
same the dataDir maxima doesn't work properly leading to
quorum issues.
- its also possible that XLV1 might be true among the disks
available, make sure to keep FileInfo based on common quorum
and skip unexpected disks with the older data format.
Also, this PR tests upgrade from older to a newer release if the
data is readable and matches the checksum.
NOTE: this is just initial work we can build on top of this to do further tests.
Replication was not working properly for encrypted
objects in single PUT object for preserving etag,
We need to make sure to preserve etag such that replication
works properly and not gets into infinite loops of copying
due to ETag mismatches.
- avoid extra lookup for 'xl.meta' since we are
definitely sure that it doesn't exist.
- use this in newMultipartUpload() as well
- also additionally do not write with O_DSYNC
to avoid loading the drives, instead create
'xl.meta' for listing operations without
O_DSYNC since these are ephemeral objects.
- do the same with newMultipartUpload() since
it gets synced when the PutObjectPart() is
attempted, we do not need to tax newMultipartUpload()
instead.
we will allow situations such as
```
a/b/1.txt
a/b
```
and
```
a/b
a/b/1.txt
```
we are going to document that this usecase is
not supported and we will never support it, if
any application does this users have to delete
the top level parent to make sure namespace is
accessible at lower level.
rest of the situations where the prefixes get
created across sets are supported as is.
This commit gathers MRF metrics from
all nodes in a cluster and return it to the caller. This will show information about the
number of objects in the MRF queues
waiting to be healed.
Each multipart upload is holding a read lock for the entire upload
duration of each part.
This makes it impossible for other parts to complete until all currently
uploading parts have released their locks.
It will also make it impossible for new parts to start as long as the
write lock is still being requested, essentially deadlocking uploads
until all that may have been granted a read lock has been completed.
Refactor to only hold the upload id lock while reading and writing
the metadata, but hold a part id lock while the part is being uploaded.
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`
In cases where a cluster is degraded, we do not uphold our consistency
guarantee and we will write fewer erasure codes and rely on healing
to recreate the missing shards.
In some cases replacing known bad disks in practice take days.
We want to change the behavior of a known degraded system to keep
the erasure code promise of the storage class for each object.
This will create the objects with the same confidence as a fully
functional cluster. The tradeoff will be that objects created
during a partial outage will take up slightly more space.
This means that when the storage class is EC:4, there should
always be written 4 parity shards, even if some disks are unavailable.
When an object is created on a set, the disks are immediately
checked. If any disks are unavailable additional parity shards
will be made for each offline disk, up to 50% of the number of disks.
We add an internal metadata field with the actual and intended
erasure code level, this can optionally be picked up later by
the scanner if we decide that data like this should be re-sharded.
However, this slice is also used for closing the writers, so close is never called on these.
Furthermore when an error is returned from a write it is now reported to the reader.
bonus: remove unused heal param from `newBitrotWriter`.
* Remove copy, now that we don't mutate.
At some places bloom filter tracker was getting
updated for `.minio.sys/tmp` bucket, there is no
reason to update bloom filters for those.
And add a missing bloom filter update for MakeBucket()
Bonus: purge unused function deleteEmptyDir()
upon errors to acquire lock context would still leak,
since the cancel would never be called. since the lock
is never acquired - proactively clear it before returning.
* lock: Always cancel the returned Get(R)Lock context
There is a leak with cancel created inside the locking mechanism. The
cancel purpose was to cancel operations such erasure get/put that are
holding non-refreshable locks.
This PR will ensure the created context.Cancel is passed to the unlock
API so it will cleanup and avoid leaks.
* locks: Avoid returning nil cancel in local lockers
Since there is no Refresh mechanism in the local locking mechanism, we
do not generate a new context or cancel. Currently, a nil cancel
function is returned but this can cause a crash. Return a dummy function
instead.
Part ETags are not available after multipart finalizes, removing this
check as not useful.
Signed-off-by: Poorna Krishnamoorthy <poorna@minio.io>
Co-authored-by: Harshavardhana <harsha@minio.io>
avoid re-read of xl.meta instead just use
the success criteria from PutObjectPart()
and check the ETag matches per Part, if
they match then the parts have been
successfully restored as is.
Signed-off-by: Harshavardhana <harsha@minio.io>
* fix: pick valid FileInfo additionally based on dataDir
historically we have always relied on modTime
to be consistent and same, we can now add additional
reference to look for the same dataDir value.
A dataDir is the same for an object at a given point in
time for a given version, let's say a `null` version
is overwritten in quorum we do not by mistake pick
up the fileInfo's incorrectly.
* make sure to not preserve fi.Data
Signed-off-by: Harshavardhana <harsha@minio.io>
This is an optimization by reducing one extra system call,
and many network operations. This reduction should increase
the performance for small file workloads.
The background healing can return NoSuchUpload error, the reason is that
healing code can return errFileNotFound with three parameters. Simplify
the code by returning exact errUploadNotFound error in multipart code.
Also ensure that a typed error is always returned whatever the number of
parameters because it is better than showing internal error.
This PR adds deadlines per Write() calls, such
that slow drives are timed-out appropriately and
the overall responsiveness for Writes() is always
up to a predefined threshold providing applications
sustained latency even if one of the drives is slow
to respond.
most of the delete calls today spend time in
a blocking operation where multiple calls need
to be recursively sent to delete the objects,
instead we can use rename operation to atomically
move the objects from the namespace to `tmp/.trash`
we can schedule deletion of objects at this
location once in 15, 30mins and we can also add
wait times between each delete operation.
this allows us to make delete's faster as well
less chattier on the drives, each server runs locally
a groutine which would clean this up regularly.
To avoid large delays in metacache cleanup, use rename
instead of recursive delete calls, renames are cheaper
move the content to minioMetaTmpBucket and then cleanup
this folder once in 24hrs instead.
If the new cache can replace an existing one, we should
let it replace since that is currently being saved anyways,
this avoids pile up of 1000's of metacache entires for
same listing calls that are not necessary to be stored
on disk.
This change moves away from a unified constructor for plaintext and encrypted
usage. NewPutObjReader is simplified for the plain-text reader use. For
encrypted reader use, WithEncryption should be called on an initialized PutObjReader.
Plaintext:
func NewPutObjReader(rawReader *hash.Reader) *PutObjReader
The hash.Reader is used to provide payload size and md5sum to the downstream
consumers. This is different from the previous version in that there is no need
to pass nil values for unused parameters.
Encrypted:
func WithEncryption(encReader *hash.Reader,
key *crypto.ObjectKey) (*PutObjReader, error)
This method sets up encrypted reader along with the key to seal the md5sum
produced by the plain-text reader (already setup when NewPutObjReader was
called).
Usage:
```
pReader := NewPutObjReader(rawReader)
// ... other object handler code goes here
// Prepare the encrypted hashed reader
pReader, err = pReader.WithEncryption(encReader, objEncKey)
```
parentDirIsObject is not using set level understanding
to check for parent objects, without this it can lead to
objects that can actually reside on a separate set as
objects and would conflict.
Current implementation requires server pools to have
same erasure stripe sizes, to facilitate same SLA
and expectations.
This PR allows server pools to be variadic, i.e they
do not have to be same erasure stripe sizes - instead
they should have SLA for parity ratio.
If the parity ratio cannot be guaranteed by the new
server pool, the deployment is rejected i.e server
pool expansion is not allowed.
This PR refactors the way we use buffers for O_DIRECT and
to re-use those buffers for messagepack reader writer.
After some extensive benchmarking found that not all objects
have this benefit, and only objects smaller than 64KiB see
this benefit overall.
Benefits are seen from almost all objects from
1KiB - 32KiB
Beyond this no objects see benefit with bulk call approach
as the latency of bytes sent over the wire v/s streaming
content directly from disk negate each other with no
remarkable benefits.
All other optimizations include reuse of msgp.Reader,
msgp.Writer using sync.Pool's for all internode calls.
The only purpose of check-dir flag in
ReadVersion is to return 404 when
an object has xl.meta but without data.
This is causing an extract call to the disk
which can be penalizing in case of busy system
where disks receive many concurrent access.
Additional cases handled
- fix address situations where healing is not
triggered on failed writes and deletes.
- consider object exists during listing when
metadata can be successfully decoded.
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.
allow directories to be replicated as well, along with
their delete markers in replication.
Bonus fix to fix bloom filter updates for directories
to be preserved.
Design: https://gist.github.com/klauspost/025c09b48ed4a1293c917cecfabdf21c
Gist of improvements:
* Cross-server caching and listing will use the same data across servers and requests.
* Lists can be arbitrarily resumed at a constant speed.
* Metadata for all files scanned is stored for streaming retrieval.
* The existing bloom filters controlled by the crawler is used for validating caches.
* Concurrent requests for the same data (or parts of it) will not spawn additional walkers.
* Listing a subdirectory of an existing recursive cache will use the cache.
* All listing operations are fully streamable so the number of objects in a bucket no
longer dictates the amount of memory.
* Listings can be handled by any server within the cluster.
* Caches are cleaned up when out of date or superseded by a more recent one.
This PR fixes a hang which occurs quite commonly at higher concurrency
by allowing following changes
- allowing lower connections in time_wait allows faster socket open's
- lower idle connection timeout to ensure that we let kernel
reclaim the time_wait connections quickly
- increase somaxconn to 4096 instead of 2048 to allow larger tcp
syn backlogs.
fixes#10413
add a hint on the disk to allow for tracking fresh disk
being healed, to allow for restartable heals, and also
use this as a way to track and remove disks.
There are more pending changes where we should move
all the disk formatting logic to backend drives, this
PR doesn't deal with this refactor instead makes it
easier to track healing in the future.
This is to ensure that Go contexts work properly, after some
interesting experiments I found that Go net/http doesn't
cancel the context when Body is non-zero and hasn't been
read till EOF.
The following gist explains this, this can lead to pile up
of go-routines on the server which will never be canceled
and will die at a really later point in time, which can
simply overwhelm the server.
https://gist.github.com/harshavardhana/c51dcfd055780eaeb71db54f9c589150
To avoid this refactor the locking such that we take locks after we
have started reading from the body and only take locks when needed.
Also, remove contextReader as it's not useful, doesn't work as expected
context is not canceled until the body reaches EOF so there is no point
in wrapping it with context and putting a `select {` on it which
can unnecessarily increase the CPU overhead.
We will still use the context to cancel the lockers etc.
Additional simplification in the locker code to avoid timers
as re-using them is a complicated ordeal avoid them in
the hot path, since locking is very common this may avoid
lots of allocations.
In `(*cacheObjects).GetObjectNInfo` copy the metadata before spawning a goroutine.
Clean up a few map[string]string copies as well, reducing allocs and simplifying the code.
Fixes#10426
MaxConnsPerHost can potentially hang a call without any
way to timeout, we do not need this setting for our proxy
and gateway implementations instead IdleConn settings are
good enough.
Also ensure to use NewRequestWithContext and make sure to
take the disks offline only for network errors.
Fixes#10304
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.
We can reduce this further in the future, but this is a good
value to keep around. With the advent of continuous healing,
we can be assured that namespace will eventually be
consistent so we are okay to avoid the necessity to
a list across all drives on all sets.
Bonus Pop()'s in parallel seem to have the potential to
wait too on large drive setups and cause more slowness
instead of gaining any performance remove it for now.
Also, implement load balanced reply for local disks,
ensuring that local disks have an affinity for
- cleanupStaleMultipartUploads()
With reduced parity our write quorum should be same
as read quorum, but code was still assuming
```
readQuorum+1
```
In all situations which is not necessary.
object KMS is configured with auto-encryption,
there were issues when using docker registry -
this has been left unnoticed for a while.
This PR fixes an issue with compatibility.
Additionally also fix the continuation-token implementation
infinite loop issue which was missed as part of #9939
Also fix the heal token to be generated as a client
facing value instead of what is remembered by the
server, this allows for the server to be stateless
regarding the token's behavior.
Bonus fix during versioning merge one of the PR was missing
the offline/online disk count fix from #9801 port it correctly
over to the master branch from release.
Additionally, add versionID support for MRF
Fixes#9910Fixes#9931
Just like GET/DELETE APIs it is possible to preserve
client supplied versionId's, of course the versionIds
have to be uuid, if an existing versionId is found
it is overwritten if no object locking policies
are found.
- PUT /bucketname/objectname?versionId=<id>
- POST /bucketname/objectname?uploads=&versionId=<id>
- PUT /bucketname/objectname?verisonId=<id> (with x-amz-copy-source)
- Implement a new xl.json 2.0.0 format to support,
this moves the entire marshaling logic to POSIX
layer, top layer always consumes a common FileInfo
construct which simplifies the metadata reads.
- Implement list object versions
- Migrate to siphash from crchash for new deployments
for object placements.
Fixes#2111
Klaus Post
Anis Elleuch
Harshavardhana
Praveen raj Mani
Krishnan Parthasarathi
Poorna Krishnamoorthy