This PR also tries to simplify the approach taken in
object-locking implementation by preferential treatment
given towards full validation.
This in-turn has fixed couple of bugs related to
how policy should have been honored when ByPassGovernance
is provided.
Simplifies code a bit, but also duplicates code intentionally
for clarity due to complex nature of object locking
implementation.
Too many deployments come up with an odd number
of hosts or drives, to facilitate even distribution
among those setups allow for odd and prime numbers
based packs.
- B2 does actually return an MD5 hash for newly uploaded objects
so we can use it to provide better compatibility with S3 client
libraries that assume the ETag is the MD5 hash such as boto.
- depends on change in blazer library.
- new behaviour is only enabled if MinIO's --compat mode is active.
- behaviour for multipart uploads is unchanged (works fine as is).
- Implement a graph algorithm to test network bandwidth from every
node to every other node
- Saturate any network bandwidth adaptively, accounting for slow
and fast network capacity
- Implement parallel drive OBD tests
- Implement a paging mechanism for OBD test to provide periodic updates to client
- Implement Sys, Process, Host, Mem OBD Infos
- total number of S3 API calls per server
- maximum wait duration for any S3 API call
This implementation is primarily meant for situations
where HDDs are not capable enough to handle the incoming
workload and there is no way to throttle the client.
This feature allows MinIO server to throttle itself
such that we do not overwhelm the HDDs.
- acquire since leader lock for all background operations
- healing, crawling and applying lifecycle policies.
- simplify lifecyle to avoid network calls, which was a
bug in implementation - we should hold a leader and
do everything from there, we have access to entire
name space.
- make listing, walking not interfere by slowing itself
down like the crawler.
- effectively use global context everywhere to ensure
proper shutdown, in cache, lifecycle, healing
- don't read `format.json` for prometheus metrics in
StorageInfo() call.
- Add conservative timeouts upto 3 minutes
for internode communication
- Add aggressive timeouts of 30 seconds
for gateway communication
Fixes#9105Fixes#8732Fixes#8881Fixes#8376Fixes#9028
This is to improve responsiveness for all
admin API operations and allowing callers
to cancel any on-going admin operations,
if they happen to be waiting too long.
canonicalize the ENVs such that we can bring these ENVs
as part of the config values, as a subsequent change.
- fix location of per bucket usage to `.minio.sys/buckets/<bucket_name>/usage-cache.bin`
- fix location of the overall usage in `json` at `.minio.sys/buckets/.usage.json`
(avoid conflicts with a bucket named `usage.json` )
- fix location of the overall usage in `msgp` at `.minio.sys/buckets/.usage.bin`
(avoid conflicts with a bucket named `usage.bin`
As an optimization of the healing, HealObjects() avoid sending an
object to the background healing subsystem when the object is
present in all disks.
However, HealObjects() should have checked the scan type, if this
deep, always pass the object to the healing subsystem.
Currently, a tree walking, needed to a list objects in a specific
set quits listing as long as it finds no entries in a disk, which
is wrong.
This affected background healing, because the latter is using
tree walk directly. If one object does not exist in the first
disk for example, it will be seemed like the object does not
exist at all and no healing work is needed.
This commit fixes the behavior.
The staleness of a lock should be determined by
the quorum number of entries returning stale,
this allows for situations when locks are held
when nodes are down - we don't accidentally
clear locks unintentionally when they are valid
and correct.
Also lock maintenance should be run by all servers,
not one server, stale locks need to be run outside
the requirement for holding distributed locks.
Thanks @klauspost for reproducing this issue
Some AWS SDKs latently rely on this value some times
to calculate the right number of parts during a parallel
GetObject request, this is feature used along with
content-range - we should support this as well.
- avoid setting last heal activity when starting self-healing
This can be confusing to users thinking that the self healing
cycle was already performed.
- add info about the next background healing round
OperationTimedout error occurs when locking
timesout, trying to acquire a lock. This
error should be returned appropriately to
the client with http status "408" (request timedout)
This translation was broken, fix it.
Bulk delete API was using cleanupObjectsBulk() which calls posix
listing and delete API to remove objects internal files in the
backend (xl.json and parts) one by one.
Add DeletePrefixes in the storage API to remove the content
of a directory in a single call.
Also use a remove goroutine for each disk to accelerate removal.
Currently the code assumed some orthogonal requirements
which led situations where when we have a setup where
we have let's say for example 168 drives, the final
set_drive_count chosen was 14. Indeed 168 drives are
divisible by 12 but this wasn't allowed due to an
unexpected requirement to have 12 to be a perfect modulo
of 14 which is not possible. This assumption was incorrect.
This PR fixes this old assumption properly, also adds
few tests and some negative tests as well. Improvements
are seen in error messages as well.
- Remove the requirement to honor storage class for deletes
- Improve `posix.DeleteFileBulk` code to Stat the volumeDir
only once per call, rather than for all object paths.
Recent modification in the code led to incorrect calculation
of offline disks.
This commit saves the endpoint list in a xlObjects then we know
the name of each disk.
lock ownership is limited to endpoints on first zone,
as we do not hold locks on other zones in an expanded
setup. current code unintentionally expired active locks
when it couldn't see ownership from the secondary zone
which leads to unexpected bugs as locking fails to work
as expected.
this PR enforces md5sum verification for following
API's to be compatible with AWS S3 spec
- PutObjectRetention
- PutObjectLegalHold
Co-authored-by: Harshavardhana <harsha@minio.io>
Allow downloading goroutine dump to help detect leaks
or overuse of goroutines.
Extensions are now type dependent.
Change `profiling` -> `profile` prefix, since that is what they are
not the abstract concept.
This is a precursor change before versioning,
removes/deprecates the requirement of remembering
partName and partETag which are not useful after
a multipart transaction has finished.
This PR reduces the overall size of the backend
JSON for large file uploads.
For a non-existent user server would return STS not initialized
```
aws --profile harsha --endpoint-url http://localhost:9000 \
sts assume-role \
--role-arn arn:xxx:xxx:xxx:xxxx \
--role-session-name anything
```
instead return an appropriate error as expected by STS API
Additionally also format the `trace` output for STS APIs
Upgrades between releases are failing due to strict
rule to avoid rolling upgrades, it is enough to
bump up APIs between versions to allow for quorum
failure and wait times. Authentication failures are
catastrophic in nature which leads to server not
be able to upgrade properly.
Fixes#9021Fixes#8968
To allow better control the cache eviction process.
Introduce MINIO_CACHE_WATERMARK_LOW and
MINIO_CACHE_WATERMARK_HIGH env. variables to specify
when to stop/start cache eviction process.
Deprecate MINIO_CACHE_EXPIRY environment variable. Cache
gc sweeps at 30 minute intervals whenever high watermark is
reached to clear least recently accessed entries in the cache
until sufficient space is cleared to reach the low watermark.
Garbage collection uses an adaptive file scoring approach based
on last access time, with greater weights assigned to larger
objects and those with more hits to find the candidates for eviction.
Thanks to @klauspost for this file scoring algorithm
Co-authored-by: Klaus Post <klauspost@minio.io>
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.
Metrics used to have its own code to calculate offline disks.
StorageInfo() was avoided because it is an expensive operation
by sending calls to all nodes.
To make metrics & server info share the same code, a new
argument `local` is added to StorageInfo() so it will only
query local disks when needed.
Metrics now calls StorageInfo() as server info handler does
but with the local flag set to false.
Co-authored-by: Praveen raj Mani <praveen@minio.io>
Co-authored-by: Harshavardhana <harsha@minio.io>
Add dummy calls which respond success when ACL's
are set to be private and fails, if user tries
to change them from their default 'private'
Some applications such as nuxeo may have an
unnecessary requirement for this operation,
we support this anyways such that don't have
to fully implement the functionality just that
we can respond with success for default ACLs
Avoid GetObjectNInfo call from cache in CopyObjectHandler
- in the case of server side copy with metadata replacement,
the reader returned from cache is never consumed, but the net
effect of GetObjectNInfo from cache layer, is cache holding a
write lock to fill the cache. Subsequent stat operation on cache in
CopyObject is not able to acquire a read lock, thus causing the hang.
Fixes#8991
we don't need to validateFormats again once we have obtained
reference format, because it is possible that at this stage
another server is doing a disk heal during startup, once
in a while due to delays we get false positives and our
server doesn't start.
Format in quorum as reference format can be assumed as valid
and we proceed further, until and unless HealFormat re-inits
the disks after a successful heal.
Also use separate port for healing tests to avoid any
conflicts with regular build testing.
Fixes#8884
RegisterNotificationTargets() cleans up all connections
that it makes to notification targets when an error occurs
during its execution.
However there is a typo in the code that makes the function to always
try to access to a nil pointer in the defer code since the function
in question will always return nil in the case of any error.
This commit fixes the typo in the code.
http.Request.ContentLength can be negative, which affects
the gateway_s3_bytes_received value in Prometheus output.
The commit only increases the value of the total received bytes
in gateway mode when r.ContentLength is greater than zero.
First step is to ensure that Path component is not decoded
by gorilla/mux to avoid routing issues while handling
certain characters while uploading through PutObject()
Delay the decoding and use PathUnescape() to escape
the `object` path component.
Thanks to @buengese and @ncw for neat test cases for us
to test with.
Fixes#8950Fixes#8647
We added support for caching and S3 related metrics in #8591. As
a continuation, it would be helpful to add support for Azure & GCS
gateway related metrics as well.
The logging subsystem was initialized under init() method in
both gateway-main.go and server-main.go which are part of
same package. This created two logging targets and hence
errors were logged twice. This PR moves the init() method
to common-main.go
Streams are returning a readcloser and returning would
decrement io count instantly, fix it.
change maxActiveIOCount to 3, meaning it will pause
crawling if 3 operations are running.
Remove the random sleep. This is running in 4 goroutines,
so mostly doing nothing.
We use the getSize latency to estimate system load,
meaning when there is little load on the system and
we get the result fast we sleep a little.
If it took a long time we have high load and release
ourselves longer.
We are sleeping inside the mutex so this affects all
goroutines doing IO.
Due to a typo in the code, a cluster was not correctly creating
`background-ops` in all disks and nodes print the following error:
minio3_1 | API: SYSTEM()
minio3_1 | Time: 19:32:45 UTC 02/06/2020
minio3_1 | DeploymentID: d67c20fa-4a1e-41f5-b319-7e3e90f425d8
minio3_1 | Error: Bucket not found: .minio.sys/background-ops
minio3_1 | 2: cmd/data-usage.go:109:cmd.runDataUsageInfo()
minio3_1 | 1: cmd/data-usage.go:56:cmd.runDataUsageInfoUpdateRoutine()
This commit fixes the typo.
Adding mutex slows down the crawler to avoid large
spikes in CPU, also add millisecond interval jitter
in calculation of disk usage to slow down the spikes
further.
This is to fix a situation where an object name incorrectly
is sent with '//' in its path heirarchy, we should reject
such object names because they may be hashed to a set where
the object might not originally belong because, this can
cause situations where once object is uploaded we cannot
delete it anymore.
Fixes#8873
This commit fixes typos in the displayed server info
w.r.t. the KMS and removes the update status.
For more information about why the update status
is removed see: PR #8943
This commit removes the `Update` functionality
from the admin API. While this is technically
a breaking change I think this will not cause
any harm because:
- The KMS admin API is not complete, yet.
At the moment only the status can be fetched.
- The `mc` integration hasn't been merged yet.
So no `mc` client could have used this API
in the past.
The `Update`/`Rewrap` status is not useful anymore.
It provided a way to migrate from one master key version
to another. However, KES does not support the concept of
key versions. Instead, key migration should be implemented
as migration from one master key to another.
Basically, the `Update` functionality has been implemented just
for Vault.
- pkg/bucket/encryption provides support for handling bucket
encryption configuration
- changes under cmd/ provide support for AES256 algorithm only
Co-Authored-By: Poorna <poornas@users.noreply.github.com>
Co-authored-by: Harshavardhana <harsha@minio.io>
XL crawling wrongly returns a zero buckets count when
there are no objects uploaded in the server yet. The reason is
data of the crawler of posix returns invalid result when all
disks has zero objects.
A simple fix is to always pick the crawling result of the first
disk but choose over the result of the disk which has the most
objects in it.
looks like 1024 buffer size is not enough in
all situations, use 8192 instead which
can satisfy all the rare situations that
may arise in base64 decoding.
multi-delete API failed with write quorum errors
under following situations
- list of files requested for delete doesn't exist
anymore can lead to quorum errors and failure
- due to usage of query param for paths, for really
long paths MinIO server rejects these requests as
malformed as unexpected.
This was reproduced with warp
The server info handler makes a http connection to other
nodes to check if they are up but does not load the custom
CAs in ~/.minio/certs/CAs.
This commit fix it.
Co-authored-by: Harshavardhana <harsha@minio.io>
JWT parsing is simplified by using a custom claim
data structure such as MapClaims{}, also writes
a custom Unmarshaller for faster unmarshalling.
- Avoid as much reflections as possible
- Provide the right types for functions as much
as possible
- Avoid strings.Join, strings.Split to reduce
allocations, rely on indexes directly.
For 'snapshot' type profiles, record a 'before' profile that can be used
as `go tool pprof -base=before ...` to compare before and after.
"Before" profiles are included in the zipped package.
[`runtime.MemProfileRate`](https://golang.org/pkg/runtime/#pkg-variables)
should not be updated while the application is running, so we set it at startup.
Co-authored-by: Harshavardhana <harsha@minio.io>
On every restart of the server, usage was being
calculated which is not useful instead wait for
sufficient time to start the crawling routine.
This PR also avoids lots of double allocations
through strings, optimizes usage of string builders
and also avoids crawling through symbolic links.
Fixes#8844
Choosing maxAllowedIOError is arbitrary and
prone to errors, when drives might be perfectly
capable of taking I/O with only few locations
return I/O error. This is a hindrance of sort
where backend filesystems like ZFS can automatically
fix and handle these scenarios.
The added problem with current approach that we
take the drive offline, making it virtually impossible
to bring it online without restart the server which
is not desirable on a busy cluster. Remove this state
such that let the backend return error appropriately
to caller and let the caller decide what to do with
the error.
return http.ErrServerClosed with proper body when
server is shutting down, allowing more context instead
of just returning '503' which doesn't mean the same
thing.
code of this form is always racy, when the
map itself is being written to as well
```
func (r Map) retMap() map[string]string {
.. lock ..
return r.internalMap
}
func (r Map) addMap(k, v string) {
.. lock ..
r.internalMap[k] = v
}
```
Anyone reading from `retMap()` is not protected
because of locking and we need to make sure
to avoid code in this manner. Always safe to
copy the map and return.
Zone abstraction of object layer was returning `nil`
incorrectly under situations where disk healing is
not required. Returning `nil` is considered as healing
successful, which leads to unexpected ReloadFormat()
peer notification calls during startup.
This PR fixes this behavior properly for zones.
Meta volumes directories, tmp/, background-ops/, etc..
undr .minio.sys are created when disks are formatted
but also when the cluster is started.
However using MakeVolBulk() is not appropriate in the
case of a user migrating from a version which does not
have .minio.sys/background-ops/. The reason is that
MakeVolBulk() exits early when an error is occured:
errVolumeExists in this case, which is expected since
some directories such as tmp/ already exist.
This commit will avoid use MakeVolBulk and use MakeVol
instead.
Also the PR will make each node creates meta volumes
in its local disks and stop relying on the first disk
since the first node could be offline.
instead perform a liveness check call to
verify if server is online and print relevant
errors.
Also introduce a StorageErr string error type
instead of errors.New() deprecate usage of
VerifyFileError, DeleteFileError for gob,
change in datastructure also requires bump in
storage REST version to v13.
Fixes#8811
Enabling the memory profiling has a significant impact on performance.
Reduce the profiling rate by 2 orders of magnitude. It is still 128x smaller than default so it should be plenty.
object lock config is enabled for a bucket.
Creating a bucket with object lock configuration
enabled does not automatically cause WORM protection
to be applied. PUT operation needs to specifically
request object locking or bucket has to have default
retention settings configured.
Fixes regression introduced in #8657
When formatting a set validate if a host failure will likely lead to data loss.
While we don't know what config will be set in the future
evaluate to our best knowledge, assuming default settings.
X-Cache sets cache status of HIT if object is
served from the disk cache, or MISS otherwise.
X-Cache-Lookup is set to HIT if object was found
in the cache even if not served (for e.g. if cache
entry was invalidated by ETag verification)
A new key was added in identity_openid recently
required explicitly for client to set the optional
value without that it would be empty, handle this
appropriately.
Fixes#8787
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
- Stop spawning store replay routines when testing the notification targets
- Properly honor the target.Close() to clean the resources used
Fixes#8707
Co-authored-by: Harshavardhana <harsha@minio.io>
In certain organizations policy claim names
can be not just 'policy' but also things like
'roles', the value of this field might also
be *string* or *[]string* support this as well
In this PR we are still not supporting multiple
policies per STS account which will require a
more comprehensive change.
Exponential backoff does not seem like a good fit for
this function since we can expect a few roundtrips on
initial startup.
This retry loop get slow pretty quickly with initial
wait being 1 second and each try being double the
wait until 30 seconds is reached.
Instead simply try 2 times per second.
This PR adds jsoniter package to replace encoding/json
in places where faster json unmarshal is necessary
whenever input JSON is large enough.
Some benchmarking comparison between jsoniter and enconding/json
benchmark old MB/s new MB/s speedup
BenchmarkParseUnmarshal/N10-4 110.02 331.17 3.01x
BenchmarkParseUnmarshal/N100-4 125.74 524.09 4.17x
BenchmarkParseUnmarshal/N500-4 131.68 542.60 4.12x
BenchmarkParseUnmarshal/N1000-4 133.93 514.88 3.84x
BenchmarkParseUnmarshal/N5000-4 122.10 415.36 3.40x
BenchmarkParseUnmarshal/N10000-4 132.13 403.90 3.06x
Fixes scenario where zones are appropriately
handled, along with supporting overriding set
count. The new fix also ensures that we handle
the various setup types properly.
Update documentation to properly indicate the
behavior.
Fixes#8750
Co-authored-by: Nitish Tiwari <nitish@minio.io>
Currently when connections to vault fail, client
perpetually retries this leads to assumptions that
the server has issues and masks the problem.
Re-purpose *crypto.Error* type to send appropriate
errors back to the client.
In existing functionality we simply return a generic
error such as "MalformedPolicy" which indicates just
a generic string "invalid resource" which is not very
meaningful when there might be multiple types of errors
during policy parsing. This PR ensures that we send
these errors back to client to indicate the actual
error, brings in two concrete types such as
- iampolicy.Error
- policy.Error
Refer #8202
minio server /data{1..4} shows an error about inability to bind a port, though
the real problem is /data{1..4} cannot be created because of the lack of
permissions.
This commit fix the behavior.
- We should declare a cluster ready even if read quorum is achieved (atleast n/2 disks are online).
- Such that, all the zones should have enough read quorum. Thus making the cluster ready for reads.
We had messy cyclical dependency problem with `mc`
due to dependencies in pkg/console, moved the pkg/console
to minio for more control and also to avoid any further
cyclical dependencies of `mc` clobbering up the
dependencies on server.
Fixes#8659
If two distinct clusters are started with different domains
along with single common domain, this situation was leading
to conflicting buckets getting created on different clusters
To avoid this do not prematurely error out if the key has no
entries, let the caller decide on which entry matches and
which entry is valid. This allows support for MINIO_DOMAIN
with one common domain, but each cluster may have their own
domains.
Fixes#8705
This is to ensure that when we have multiple tenants
deployed all sharing the same etcd for global bucket
should avoid listing each others buckets, this leads
to information leak which should be avoided unless
etcd is not namespaced for IAM assets in which case
it can be assumed that its a federated setup.
Federated setup and namespaced IAM assets on etcd
is not supported since namespacing is only useful
when you wish to separate the tenants as isolated
instances of MinIO.
This PR allows a new type of behavior, primarily
driven by the usecase of m3(mkube) multi-tenant
deployments with global bucket support.
Harshavardhana
Krishna Srinivas
ebozduman
poornas
Bala FA
Sidhartha Mani
Ingmar Runge
Anis Elleuch
Nitish Tiwari
Stephen N
Klaus Post
kannappanr
Kody A Kantor
Praveen raj Mani
astorath
Andreas Auernhammer
Krishnan Parthasarathi
Ashish Kumar Sinha
Aditya Manthramurthy
Forest Lovewood
George Xie