GetActualSize() was heavily relying on o.Parts()
to be non-empty to figure out if the object is multipart or not,
However, we have many indicators of whether an object is multipart
or not.
Blindly assuming that o.Parts == nil is not a multipart, is an
incorrect expectation instead, multipart must be obtained via
- Stored metadata value indicating this is a multipart encrypted object.
- Rely on <meta>-actual-size metadata to get the object's actual size.
This value is preserved for additional reasons such as these.
- ETag != 32 length
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.
This PR also increases per node bpool memory from 1024 entries
to 2048 entries; along with that, it also moves the byte pool
centrally instead of being per pool.
there is potential for danglingWrites when quorum failed, where
only some drives took a successful write, generally this is left
to the healing routine to pick it up. However it is better that
we delete it right away to avoid potential for quorum issues on
version signature when there are many versions of an object.
AWS S3 closes keep-alive connections frequently
leading to frivolous logs filling up the MinIO
logs when the transition tier is an AWS S3 bucket.
Ignore such transient errors, let MinIO retry
it when it can.
Optionally allows customers to enable
- Enable an external cache to catch GET/HEAD responses
- Enable skipping disks that are slow to respond in GET/HEAD
when we have already achieved a quorum
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.
Currently if the object does not exist in quorum disks of an erasure
set, the dangling code is never called because the returned error will
be errFileNotFound or errFileVersionNotFound;
With this commit, when errFileNotFound or errFileVersionNotFound is
returning when trying to calculate the quorum of a given object, the
code checks if a disk returned nil, which means a stale object exists in
that disk, that will trigger deleteIfDangling() function
if erasure upgrade is needed rely on the in-memory
values, instead of performing a "DiskInfo()" call.
https://brendangregg.com/blog/2016-09-03/sudden-disk-busy.html
for HDDs these are problematic, lets avoid this because
there is no value in "being" absolutely strict here
in terms of parity. We are okay to increase parity
as we see based on the in-memory online/offline ratio.
without this the rename2() can rename the previous dataDir
causing issues for different versions of the object, only
latest version is preserved due to this bug.
Added healing code to ensure recovery of such content.
Errors such as
```
returned an error (context deadline exceeded) (*fmt.wrapError)
```
```
(msgp: too few bytes left to read object) (*fmt.wrapError)
```
objects with 10,000 parts and many of them can
cause a large memory spike which can potentially
lead to OOM due to lack of GC.
with previous PR reducing the memory usage significantly
in #17963, this PR reduces this further by 80% under
repeated calls.
Scanner sub-system has no use for the slice of Parts(),
it is better left empty.
```
benchmark old ns/op new ns/op delta
BenchmarkToFileInfo/ToFileInfo-8 295658 188143 -36.36%
benchmark old allocs new allocs delta
BenchmarkToFileInfo/ToFileInfo-8 61 60 -1.64%
benchmark old bytes new bytes delta
BenchmarkToFileInfo/ToFileInfo-8 1097210 227255 -79.29%
```
- this PR avoids sending a large ChecksumInfo slice
when its not needed
- also for a file with XLV2 format there is no reason
to allocate Checksum slice while reading
.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.
Bonus:
- avoid calling DiskInfo() calls when missing blocks
instead heal the object using MRF operation.
- change the max_sleep to 250ms beyond that we will
not stop healing.
on unversioned buckets its possible that 0-byte objects
might lose quorum on flaky systems, allow them to be same
as DELETE markers. Since practically speak they have no
content.
Optimize DeleteObject API to avoid extra
GetObjectInfo call on the replicating side.
For receiving side, it is just a regular
DeleteObject call.
Bonus: Fix a corner case where version purged is
absent on target (either due to replication not yet
complete or target version already deleted in a
one-way replication or when replication was disabled).
In such cases, mark version purge complete.
* 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
there is a possibility that slow drives can actually add latency
to the overall call, leading to a large spike in latency.
this can happen if there are other parallel listObjects()
calls to the same drive, in-turn causing each other to sort
of serialize.
this potentially improves performance and makes PutObject()
also non-blocking.
on "unversioned" buckets there are situations
when successive concurrent I/O can lead to
an inconsistent state() with mtime while the
etag might be the same for the object on disk.
in such a scenario it is possible for us to
allow reading of the object since etag matches
and if etag matches we are guaranteed that we
have enough copies the object will be readable
and same.
This PR allows fallback in such scenarios.
fixes an issue under bucket replication could cause
ETags for replicated SSE-S3 single part PUT objects,
to fail as we would attempt a decryption while listing,
or stat() operation.
Removes the bloom filter since it has so limited usability, often gets saturated anyway and adds a bunch of complexity to the scanner.
Also removes a tiny bit of CPU by each write operation.
Move to using `xl.meta` data structure to keep temporary partInfo,
this allows for a future change where we move to different parts to
different drives.
PUT shall only proceed if pre-conditions are met, the new
code uses
- x-minio-source-mtime
- x-minio-source-etag
to verify if the object indeed needs to be replicated
or not, allowing us to avoid StatObject() call.
This PR is a continuation of the previous change instead
of returning an error, instead trigger a spot heal on the
'xl.meta' and return only after the healing is complete.
This allows for future GETs on the same resource to be
consistent for any version of the object.
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
The bottom line is delete markers are a nuisance,
most applications are not version aware and this
has simply complicated the version management.
AWS S3 gave an unnecessary complication overhead
for customers, they need to now manage these
markers by applying ILM settings and clean
them up on a regular basis.
To make matters worse all these delete markers
get replicated as well in a replicated setup,
requiring two ILM settings on each site.
This PR is an attempt to address this inferior
implementation by deviating MinIO towards an
idempotent delete marker implementation i.e
MinIO will never create any more than single
consecutive delete markers.
This significantly reduces operational overhead
by making versioning more useful for real data.
This is an S3 spec deviation for pragmatic reasons.
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.
