Bonus: if runs have PUT higher then capture it anyways
to display an unexpected result, which provides a way
to understand what might be slowing things down on the
system.
For example on a Data24 WDC setup it is clearly visible
there is a bug in the hardware.
```
./mc admin speedtest wdc/
⠧ Running speedtest (With 64 MiB object size, 32 concurrency) PUT: 31 GiB/s GET: 24 GiB/s
⠹ Running speedtest (With 64 MiB object size, 48 concurrency) PUT: 38 GiB/s GET: 24 GiB/s
MinIO 2021-11-04T06:08:33Z, 6 servers, 48 drives
PUT: 38 GiB/s, 605 objs/s
GET: 24 GiB/s, 383 objs/s
```
Reads are almost 14GiB/sec slower than Writes which
is practically not possible.
This reverts commit 091a7ae359.
- Ensure all actions accessing storage lock properly.
- Behavior change: policies can be deleted only when they
are not associated with any active credentials.
Also adds fix for accidental canned policy removal that was present in the
reverted version of the change.
Windows users often click on the binary without
knowing MinIO is a command-line tool and should be
run from a terminal. Throw a message to guide them
on what to do.
Co-authored-by: Klaus Post <klauspost@gmail.com>
Borrowed idea from Go's usage of this
optimization for ReadFrom() on client
side, we should re-use the 32k buffers
io.Copy() allocates for generic copy
from a reader to writer.
the performance increase for reads for
really tiny objects is at this range
after this change.
> * Fastest: +7.89% (+1.3 MiB/s) throughput, +7.89% (+1308.1) obj/s
- Ensure all actions accessing storage lock properly.
- Behavior change: policies can be deleted only when they
are not associated with any active credentials.
- The race happens with a goroutine that refreshes IAM cache data from storage.
- It could lead to deleted users re-appearing as valid live credentials.
- This change also causes CI to run tests without a race flag (in addition to
running it with).
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.
- combine similar looking functionalities into single
handlers, and remove unnecessary proxying of the
requests at handler layer.
- remove bucket forwarding handler as part of default setup
add it only if bucket federation is enabled.
Improvements observed for 1kiB object reads.
```
-------------------
Operation: GET
Operations: 4538555 -> 4595804
* Average: +1.26% (+0.2 MiB/s) throughput, +1.26% (+190.2) obj/s
* Fastest: +4.67% (+0.7 MiB/s) throughput, +4.67% (+739.8) obj/s
* 50% Median: +1.15% (+0.2 MiB/s) throughput, +1.15% (+173.9) obj/s
```
Removes RLock/RUnlock for updating metadata,
since we already take a write lock to update
metadata, this change removes reading of xl.meta
as well as an additional lock, the performance gain
should increase 3x theoretically for
- PutObjectRetention
- PutObjectLegalHold
This optimization is mainly for Veeam like
workloads that require a certain level of iops
from these API calls, we were losing iops.
read/writers are not concurrent in handlers
and self contained - no need to use atomics on
them.
avoids unnecessary contentions where it's not
required.
Logger targets were not race protected against concurrent updates from for example `HTTPConsoleLoggerSys`.
Restrict direct access to targets and make slices immutable so a returned slice can be processed safely without locks.
As we use etcd's watch interface, we do not need the
network notifications as they are no-ops anyway.
Bonus: Remove globalEtcdClient global usage in IAM
IAMSys is a higher-level object, that should not be called by the lower-level
storage API interface for IAM. This is to prepare for further improvements in
IAM code.
etcd operations, get/put/delete, should be logged when failed
with errors other than not found error. It will make it easier to
see connections issues from MinIO to etcd.
Refresh was doing a linear scan of all locked resources. This was adding
up to significant delays in locking on high load systems with long
running requests.
Add a secondary index for O(log(n)) UID -> resource lookups.
Multiple resources are stored in consecutive strings.
Bonus fixes:
* On multiple Unlock entries unlock the write locks we can.
* Fix `expireOldLocks` skipping checks on entry after expiring one.
* Return fast on canTakeUnlock/canTakeLock.
* Prealloc some places.
We do not reliably know the length of compressed data, including headers.
Request until the end-of-stream. Results will still be properly truncated.
Fixes#13441
Testing with `mc sql --compression BZIP2 --csv-input "rd=\n,fh=USE,fd=;" --query="select COUNT(*) from S3Object" local2/testbucket/nyc-taxi-data-10M.csv.bz2`
Before 96.98s, after 10.79s. Uses about 70% CPU while running.
offset+length should match the Size() of the individual parts
return 'errFileCorrupt' otherwise, to trigger healing of the individual
parts do not error out prematurely when healing such bitrot's upon
successful parts being written to the client.
another issue this PR fixes is to not return and error to
the client if we have just triggered a heal on a specific
part of the object, instead continue to read all the content
and let the heal happen asynchronously later.
Looks like policy restriction was not working properly
for normal users when they are not svc or STS accounts.
- svc accounts are now properly fixed to get
right permissions when its inherited, so
we do not have to set 'owner = true'
- sts accounts have always been using right
permissions, do not need an explicit lookup
- regular users always have proper policy mapping
- avoids relying in listQuorum from the underlying listObjects()
and potentially missing entries if any.
- avoid the entire merging logic etc, listing raw set by set
and loading whatever is found is cleaner when dealing with
a large cluster for IAM metadata.
