This type of code is not necessary, read's of all
metadata content at `.minio.sys/config` automatically
triggers healing when necessary in the GetObjectNInfo()
call-path.
Having this code is not useful and this also adds to
the overall startup time of MinIO when there are lots
of users and policies.
The main goal of this PR is to solve the situation where disks stop
responding to operations. This generally causes an FD build-up and
eventually will crash the server.
This adds detection of hung disks, where calls on disk get stuck.
We add functionality to `xlStorageDiskIDCheck` where it keeps
track of the number of concurrent requests on a given disk.
A total number of 100 operations are allowed. If this limit is reached
we will block (but not reject) new requests, but we will monitor the
state of the disk.
If no requests have been completed or updated within a 15-second
window, we mark the disk as offline. Requests that are blocked will be
unblocked and return an error as "faulty disk".
New requests will be rejected until the disk is marked OK again.
Once a disk has been marked faulty, a check will run every 5 seconds that
will attempt to write and read back a file. As long as this fails the disk will
remain faulty.
To prevent lots of long-running requests to mark the disk faulty we
implement a callback feature that allows updating the status as parts
of these operations are running.
We add a reader and writer wrapper that will update the status of each
successful read/write operation. This should allow fine enough granularity
that a slow, but still operational disk will not reach 15 seconds where
50 operations have not progressed.
Note that errors themselves are not enough to mark a disk faulty.
A nil (or io.EOF) error will mark a disk as "good".
* Make concurrent disk setting configurable via `_MINIO_DISK_MAX_CONCURRENT`.
* de-couple IsOnline() from disk health tracker
The purpose of IsOnline() is to ensure that we
reconnect the drive only when the "drive" was
- disconnected from network we need to validate
if the drive is "correct" and is the same drive
which belongs to this server.
- drive was replaced we have to format it - we
support hot swapping of the drives.
IsOnline() is not meant for taking the drive offline
when it is hung, it is not useful we can let the
drive be online instead "return" errors for relevant
calls.
* return errFaultyDisk for DiskInfo() call
Co-authored-by: Harshavardhana <harsha@minio.io>
Possible future Improvements:
* Unify the REST server and local xlStorageDiskIDCheck. This would also improve stats significantly.
* Allow reads/writes to be aborted by the context.
* Add usage stats, concurrent count, blocked operations, etc.
Data usage does not always contain tiering info even if the data usage
information is valid. Avoid a crash in that case.
(e.g. the scanner scanned the namespace, the user enables tiering,
prometheus scrapes the server before the scanner gets a chance to
update the data usage with new tiering information)
Healing decisions would align with skipped folder counters. This can lead to files
never being selected for heal checks on "clean" paths.
Use different hashing methods and take objectHealProbDiv into account when
calculating the cycle.
Found by @vadmeste
This is a side-affect of the optimization done in PR #13544 which
causes a certain type of delete operations on given object versions
can cause lastVersion indication to be skipped, which leads to
an `xl.meta` where Versions[] slice is empty while the entire
file is intact by itself.
This PR tries to ensure that such files are visible and deletable
by regular means of listing as null 'delete-marker' and also
avoid the situation where this potential issue might arise.
When scanning using normal mode, HealObject() can report an
error saying that it found a corrupted part. This doesn't have
when HealObject() is called with bitrot scan flag. However, when
this happens, we can still restart HealObject() with the bitrot scan.
This is also important because this means the scanner and the
new disks healer will not be able to heal an object that doesn't
exist in a specific disk and has corruption in another disk.
Also without this PR, mc admin heal command without bitrot will report
an error.
This commit removes some duplicate code that
converts KES API errors.
This code was added since KES `0.18.0` changed
some exported API errors. However, the KES SDK
handles this error conversion itself.
Therefore, it is not necessary to duplicate this
behavior in MinIO.
See: 21555fa624/error.go (L94)
Signed-off-by: Andreas Auernhammer <hi@aead.dev>
- Updating KES dependency to v.0.18.0
- Fixing incompatibility issue when checking for errors during KES key creation
Signed-off-by: Lenin Alevski <alevsk.8772@gmail.com>
In a distributed setup, a DiskInfo REST call to an unformatted disk
returns an error with no disk information, such as the disk endpoint
URL, which is unexpected.
metadata headers can have headers without values
as per AWS S3 spec however, we need to skip some
headers that do not have values that potentially
can have empty values set.
small setups do not return appropriate errors when speedtest
cannot run on small tiny setups, allow the tests to fail
appropriately more pro-actively.
many users bring toy setups, this PR simply returns an error
in such situations.
healing disks take active I/O it is possible
that deleted objects might stay in .trash
folder for a really long time until the drive
is fully healed.
this PR changes it such that we are making sure
we purge the active content written to these
disks as well.
- speedtest logs calls that were canceled
spuriously, in situations where it should
be ignored.
- all errors of interest are always sent back
to the client there is no need to log them
on the server console.
- PUT failures should negate the increments
such that GET is not attempted on unsuccessful
calls.
- do not attempt MRF on speedtest objects.
In the testing mode, reformatting disks will fail because the healing
code will complain if one disk is in root mode. This commit will
automatically set all disks as non-root if MINIO_CI_CD is set.
Currently, when applying any dynamic config, the system reloads and
re-applies the config of all the dynamic sub-systems.
This PR refactors the code in such a way that changing config of a given
dynamic sub-system will work on only that sub-system.
An onlineDisk means its a valid disk but it may be a
re-connected disk, this PR verifies that based on LastConn()
to only trigger MRF. Current code would again re-load the
disk 'format.json' which is not necessary and perhaps an
unnecessary call.
A potential side affect of this is closing perfectly online
disks and getting re-replaced by reloading 'format.json'.
This PR tries to avoid this situation by making sure MRF
is triggered but not reloading 'format.json' because of MRF.
Only the first `listAndHeal` would ever be able to write on errCh, blocking all others infinitely.
Instead read all errors but return the first non-nil, if any.
The intention appears to be that this should cancel on any error,
so that part is kept.
Regression from #13990
When more than one gateway reads and writes from the same mount point
and there is a load balancer pointing to those gateways. Each gateway
will try to create its own temporary append file but fails to clear it later
when not needed.
This commit creates a routine that checks all upload IDs saved in
multipart directory and remove any stale entry with the same upload id
in the memory and in the temporary background append folder as well.
Enabled with `mc admin config set alias/ api gzip_objects=on`
Standard filtering applies (1K response minimum, not compressed content
type, not range request, gzip accepted by client).
The current code considers a pool with all root disks to be as part
of a testing environment even if there are other pools with mounted
disks. This will result to illegitimate writing in root disks.
Fix this by simplifing the logic: require MINIO_CI_CD in order to skip
root disk check.
MinIO configuration is loaded after the initialization of the server
handlers, which will miss the initialization of the bucket forwarder
handler.
Though the federation is deprecated, let's fix this for the time being.
S3 spec returns x-amz-restore header in HEAD/GET object with the
following format:
```
x-amz-restore: ongoing-request="false", expiry-date="Fri, 21 Dec 2012
00:00:00 GMT"
```
This commit adds quotes as the current code does not support it. It will
also supports the old format saved in the disk (in xl.meta) for backward
compatibility.
A regression removed support of federation in the gateway mode.
Enable it again.
Federation is deprecated for a while but let's fix this for the time being.
Deleting bulk objects had an issue since the relevant versionID
is not passed through the layers to ensure that the dangling
object purge actually works cleanly.
This is a continuation of quorum related error returned by
multi-object delete API from #14248
This PR ensures that we pass down correct information as
well as extend the scope of dangling object detection.
When setting a config of a particular sub-system, validate the existing
config and notification targets of only that sub-system, so that
existing errors related to one sub-system (e.g. notification target
offline) do not result in errors for other sub-systems.
Some users running MinIO claim that their system became slow. One
way to investigate is to look at this Prometheus history of the number of
the requests reaching the server. The existing current S3 requests metric
is not enough because it can increase of the system really becomes slow,
due to disk issues for example.
startup speed-up, currently getFormatErasureInQuorum()
would spend up to 2-3secs when there are 3000+ drives
for example in a setup, simplify this implementation
to use drive counts.
DeleteMarkers do not have a default quorum, i.e it is possible that
DeleteMarkers were created with n/2+1 quorum as well to make sure
that we satisfy situations such as those we need to make sure delete
markers only expect n/2 read quorum.
Additionally we should also look at additional metadata on the
actual objects that might have been "erasure" upgraded with new
parity when disks are down.
In such a scenario do not default to the standard storage class
parity, instead use the parityBlocks present on the FileInfo to
ensure that we are dealing with the correct quorum for READs and
DELETEs.
Retry listings, when no next marker is returned and the result isn't truncated.
This can happen when an object is queued, but no info can be fetched.
Fixes#14190
The healing code repeatedly tries to heal a root disk when it is empty
the reason is that connectEndpoint() returns errUnformattedDisk even
if the disk is a root disk. Changing that to returning another error
will avoid queueing the disk to the healing code in each connect disks
iteration.
some upgraded objects might not get listed due
to different quorum ratios across objects.
make sure to list all objects that satisfy the
maximum possible quorum.
This change allows the MinIO server to lookup users in different directory
sub-trees by allowing specification of multiple search bases separated by
semicolons.
This PR removes an unnecessary state that gets
passed around for DiskIDs, which is not necessary
since each disk exactly knows which pool and which
set it belongs to on a running system.
Currently cached DiskId's won't work properly
because it always ends up skipping offline disks
and never runs healing when disks are offline, as
it expects all the cached diskIDs to be present
always. This also sort of made things in-flexible
in terms perhaps a new diskID for `format.json`.
(however this is not a big issue)
This is an unnecessary requirement that healing
via scanner needs all drives to be online, instead
healing should trigger even when partial nodes
and drives are available this ensures that we
keep the SLA in-tact on the objects when disks
are offline for a prolonged period of time.
Wrong resource is being fetched, since idx is incremented, but mapID is reused.
Regression caused by #13454 - that part didn't optimize anything anyway.
Publish storage functions latency to help compare the performance
of different disks in a single deployment.
e.g.:
```
minio_node_disk_latency_us{api="storage.WalkDir",disk="/tmp/xl/1",server="localhost:9001"} 226
minio_node_disk_latency_us{api="storage.WalkDir",disk="/tmp/xl/2",server="localhost:9002"} 1180
minio_node_disk_latency_us{api="storage.WalkDir",disk="/tmp/xl/3",server="localhost:9003"} 1183
minio_node_disk_latency_us{api="storage.WalkDir",disk="/tmp/xl/4",server="localhost:9004"} 1625
```
- create internal erasure volumes only if the disk is unformatted
- return a copy of format data in xlStorage.ReadAll
- parse env vars only once, to be re-used by xl-storage
This speed-up is intended for faster startup times
for almost all MinIO operations. Changes here are
- Drives are not re-read for 'format.json' on a regular
basis once read during init is remembered and refreshed
at 5 second intervals.
- Do not do O_DIRECT tests on drives with existing 'format.json'
only fresh setups need this check.
- Parallelize initializing erasureSets for multiple sets.
- Avoid re-reading format.json when migrating 'format.json'
from really old V1->V2->V3
- Keep a copy of local drives for any given server in memory
for a quick lookup.
this helps in caching the resolved values early on, avoids
causing further resolution for individual nodes when
object layer comes online.
this can speed up our startup time during, upgrades etc by
an order of magnitude.
additional changes in connectLoadInitFormats() and parallelize
all calls that might be potentially blocking.
- Site replication was missing replicating users,
groups when an empty site was added.
- Add site replication for groups and users when they
are disabled and enabled.
- Add support for replicating bucket quota config.
When calculating signatures empty part ETags were not discarded, leading
to a different signature compared to freshly created ones.
This would mean that after a heal signature of the healed metadata would be
different. Fixing the calculation of signature will make these consistent.
Furthermore when inconsistent entries, with zero version ID, with the same
mod times but different signatures, the one with the lowest signature would
be picked for quorum check. Since this is 50/50, we fall back to a simple
quorum count on all signatures.
Each of these fixes by themselves will lead to quorum. Tests were added
for regressions and expected outcomes.
When the replication rule is based on tag matches, the replication process
should pick up targets matching the tags specified in the replication
rule.
Fixing regression due to #12880
repeated reads on single large objects in HPC like
workloads, need the following option to disable
O_DIRECT for a more effective usage of the kernel
page-cache.
However this optional should be used in very specific
situations only, and shouldn't be enabled on all
servers.
NVMe servers benefit always from keeping O_DIRECT on.
map labels might have been referenced else, this
can lead to concurrent access at lower layers.
avoid this by copying the information while
concurrently serving the metrics.
do not allow mutation to pool command line when there are
unfinished decommissions in place, disallow such scenarios
to avoid user mistakes.
also add testcases to cover all relevant scenarios.
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.
When deleting multiple versions it "gives" up with an errFileVersionNotFound if
a version cannot be found. This effectively skips deleting other versions
sent in the same request.
This can happen on inconsistent objects. We should ignore errFileVersionNotFound
and continue with others.
We already ignore these at the caller level, this PR is continuation of 54a9877
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.
Large clusters with multiple sets, or multi-pool setups at times might
fail and report unexpected "file not found" errors. This can become
a problem during startup sequence when some files need to be created
at multiple locations.
- This PR ensures that we nil the erasure writers such that they
are skipped in RenameData() call.
- RenameData() doesn't need to "Access()" calls for `.minio.sys`
folders they always exist.
- Make sure PutObject() never returns ObjectNotFound{} for any
errors, make sure it always returns "WriteQuorum" when renameData()
fails with ObjectNotFound{}. Return appropriate errors for all
other cases.
Currently tag removal leaves replication state as `PENDING`
because the `HEAD` api returns just a tag count but not the
actual tags, and this is treated as a no-op
```
λ mc admin decommission start alias/ http://minio{1...2}/data{1...4}
```
```
λ mc admin decommission status alias/
┌─────┬─────────────────────────────────┬──────────────────────────────────┬────────┐
│ ID │ Pools │ Capacity │ Status │
│ 1st │ http://minio{1...2}/data{1...4} │ 439 GiB (used) / 561 GiB (total) │ Active │
│ 2nd │ http://minio{3...4}/data{1...4} │ 329 GiB (used) / 421 GiB (total) │ Active │
└─────┴─────────────────────────────────┴──────────────────────────────────┴────────┘
```
```
λ mc admin decommission status alias/ http://minio{1...2}/data{1...4}
Progress: ===================> [1GiB/sec] [15%] [4TiB/50TiB]
Time Remaining: 4 hours (started 3 hours ago)
```
```
λ mc admin decommission status alias/ http://minio{1...2}/data{1...4}
ERROR: This pool is not scheduled for decommissioning currently.
```
```
λ mc admin decommission cancel alias/
┌─────┬─────────────────────────────────┬──────────────────────────────────┬──────────┐
│ ID │ Pools │ Capacity │ Status │
│ 1st │ http://minio{1...2}/data{1...4} │ 439 GiB (used) / 561 GiB (total) │ Draining │
└─────┴─────────────────────────────────┴──────────────────────────────────┴──────────┘
```
> NOTE: Canceled decommission will not make the pool active again, since we might have
> Potentially partial duplicate content on the other pools, to avoid this scenario be
> very sure to start decommissioning as a planned activity.
```
λ mc admin decommission cancel alias/ http://minio{1...2}/data{1...4}
┌─────┬─────────────────────────────────┬──────────────────────────────────┬────────────────────┐
│ ID │ Pools │ Capacity │ Status │
│ 1st │ http://minio{1...2}/data{1...4} │ 439 GiB (used) / 561 GiB (total) │ Draining(Canceled) │
└─────┴─────────────────────────────────┴──────────────────────────────────┴────────────────────┘
```
In a multi-pool setup when disks are coming up, or in a single pool
setup let's say with 100's of erasure sets with a slow network.
It's possible when healing is attempted on `.minio.sys/config`
folder, it can lead to healing unexpectedly deleting some policy
files as dangling due to a mistake in understanding when `isObjectDangling`
is considered to be 'true'.
This issue happened in commit 30135eed86
when we assumed the validMeta with empty ErasureInfo is considered
to be fully dangling. This implementation issue gets exposed when
the server is starting up.
This is most easily seen with multiple-pool setups because of the
disconnected fashion pools that come up. The decision to purge the
object as dangling is taken incorrectly prior to the correct state
being achieved on each pool, when the corresponding drive let's say
returns 'errDiskNotFound', a 'delete' is triggered. At this point,
the 'drive' comes online because this is part of the startup sequence
as drives can come online lazily.
This kind of situation exists because we allow (totalDisks/2) number
of drives to be online when the server is being restarted.
Implementation made an incorrect assumption here leading to policies
getting deleted.
Added tests to capture the implementation requirements.
- This allows site-replication to be configured when using OpenID or the
internal IDentity Provider.
- Internal IDP IAM users and groups will now be replicated to all members of the
set of replicated sites.
- When using OpenID as the external identity provider, STS and service accounts
are replicated.
- Currently this change dis-allows root service accounts from being
replicated (TODO: discuss security implications).
It is possible that GetLock() call remembers a previously
failed releaseAll() when there are networking issues, now
this state can have potential side effects.
This PR tries to avoid this side affect by making sure
to initialize NewNSLock() for each GetLock() attempts
made to avoid any prior state in the memory that can
interfere with the new lock grants.
The current usage of assuming `default` parity of `4` is not correct
for all objects stored on MinIO, objects in .minio.sys have maximum
parity, healing won't trigger on these objects due to incorrect
verification of quorum.
The AddUser() API endpoint was accepting a policy field.
This API is used to update a user's secret key and account
status, and allows a regular user to update their own secret key.
The policy update is also applied though does not appear to
be used by any existing client-side functionality.
This fix changes the accepted request body type and removes
the ability to apply policy changes as that is possible via the
policy set API.
NOTE: Changing passwords can be disabled as a workaround
for this issue by adding an explicit "Deny" rule to disable the API
for users.
This PR is an attempt to make this configurable
as not all situations have same level of tolerable
delta, i.e disks are replaced days apart or even
hours.
There is also a possibility that nodes have drifted
in time, when NTP is not configured on the system.
data shards were wrong due to a healing bug
reported in #13803 mainly with unaligned object
sizes.
This PR is an attempt to automatically avoid
these shards, with available information about
the `xl.meta` and actually disk mtime.
- When using MinIO's internal IDP, STS credential permissions did not check the
groups of a user.
- Also fix bug in policy checking in AccountInfo call
Also log all the missed events and logs instead of silently
swallowing the events.
Bonus: Extend the logger webhook to support mTLS
similar to audit webhook target.
- r.ulock was not locked when r.UsageCache was being modified
Bonus:
- simplify code by removing some unnecessary clone methods - we can
do this because go arrays are values (not pointers/references) that are
automatically copied on assignment.
- remove some unnecessary map allocation calls
data-structures were repeatedly initialized
this causes GC pressure, instead re-use the
collectors.
Initialize collectors in `init()`, also make
sure to honor the cache semantics for performance
requirements.
Avoid a global map and a global lock for metrics
lookup instead let them all be lock-free unless
the cache is being invalidated.
When STS credentials are created for a user, a unique (hopefully stable) parent
user value exists for the credential, which corresponds to the user for whom the
credentials are created. The access policy is mapped to this parent-user and is
persisted. This helps ensure that all STS credentials of a user have the same
policy assignment at all times.
Before this change, for an OIDC STS credential, when the policy claim changes in
the provider (when not using RoleARNs), the change would not take effect on
existing credentials, but only on new ones.
To support existing STS credentials without parent-user policy mappings, we
lookup the policy in the policy claim value. This behavior should be deprecated
when such support is no longer required, as it can still lead to stale
policy mappings.
Additionally this change also simplifies the implementation for all non-RoleARN
STS credentials. Specifically, for AssumeRole (internal IDP) STS credentials,
policies are picked up from the parent user's policies; for
AssumeRoleWithCertificate STS credentials, policies are picked up from the
parent user mapping created when the STS credential is generated.
AssumeRoleWithLDAP already picks up policies mapped to the virtual parent user.
A corner case can occur where the delete-marker was propagated
but the metadata could not be updated on the primary. Sending
a RemoveObject call with the Delete marker version would end
up permanently deleting the version on target. Instead, perform
a Stat on the delete-marker version on target and redo replication
only if the delete-marker is missing on target.
After the introduction of Refresh logic in locks, the data scanner can
quit when the data scanner lock is not able to get refreshed. In that
case, the context of the data scanner will get canceled and
runDataScanner() will quit. Another server would pick the scanning
routine but after some time, all nodes can just have all scanning
routine aborted, as described above.
This fix will just run the data scanner in a loop.
- Allow proper SRError to be propagated to
handlers and converted appropriately.
- Make sure to enable object locking on buckets
when requested in MakeBucketHook.
- When DNSConfig is enabled attempt to delete it
first before deleting buckets locally.
- Rename MaxNoncurrentVersions tag to NewerNoncurrentVersions
Note: We apply overlapping NewerNoncurrentVersions rules such that
we honor the highest among applicable limits. e.g if 2 overlapping rules
are configured with 2 and 3 noncurrent versions to be retained, we
will retain 3.
- Expire newer noncurrent versions after noncurrent days
- MinIO extension: allow noncurrent days to be zero, allowing expiry
of noncurrent version as soon as more than configured
NewerNoncurrentVersions are present.
- Allow NewerNoncurrentVersions rules on object-locked buckets
- No x-amz-expiration when NewerNoncurrentVersions configured
- ComputeAction should skip rules with NewerNoncurrentVersions > 0
- Add unit tests for lifecycle.ComputeAction
- Support lifecycle rules with MaxNoncurrentVersions
- Extend ExpectedExpiryTime to work with zero days
- Fix all-time comparisons to be relative to UTC
- This introduces a new admin API with a query parameter (v=2) to return a
response with the timestamps
- Older API still works for compatibility/smooth transition in console
- allow any regular user to change their own password
- allow STS credentials to create users if permissions allow
Bonus: do not allow changes to sts/service account credentials (via add user API)
ListObjects() should never list a delete-marked folder
if latest is delete marker and delimiter is not provided.
ListObjectVersions() should list a delete-marked folder
even if latest is delete marker and delimiter is not
provided.
Enhance further versioning listing on the buckets
request.Form uses less memory allocation and avoids gorilla mux matching
with weird characters in parameters such as '\n'
- Remove Queries() to avoid matching
- Ensure r.ParseForm is called to populate fields
- Add a unit test for object names with '\n'
delete marked objects should not be considered
for listing when listing is delimited, this issue
as introduced in PR #13804 which was mainly to
address listing of directories in listing when
delimited.
This PR fixes this properly and adds tests to
ensure that we behave in accordance with how
an S3 API behaves for ListObjects() without
versions.
Save part.1 for writebacks in a separate folder
and move it to cache dir atomically while saving
the cache metadata. This is to avoid GC mistaking
part.1 as orphaned cache entries and purging them.
This PR also fixes object size being overwritten during
retries for write-back mode.
- deleting policies was deleting all LDAP
user mapping, this was a regression introduced
in #13567
- deleting of policies is properly sent across
all sites.
- remove unexpected errors instead embed the real
errors as part of the 500 error response.
- deleteBucket() should be called for cleanup
if client abruptly disconnects
- out of disk errors should be sent to client
properly and also cancel the calls
- limit concurrency to available MAXPROCS not
32 for auto-tuned setup, if procs are beyond
32 then continue normally. this is to handle
smaller setups.
fixes#13834
Return errors when untar fails at once.
Current error handling was quite a mess. Errors are written
to the stream, but processing continues.
Instead, return errors when they occur and transform
internal errors to bad request errors, since it is likely a
problem with the input.
Fixes#13832
Sometimes, we see an error message like "Server expects 'storage' API
version 'v41', instead found 'v41'" shows a more generic error message
with the path of the REST call.
The earlier approach of using a license token for
communicating with SUBNET is being replaced
with a simpler mechanism of API keys. Unlike the
license which is a JWT token, these API keys will
be simple UUID tokens and don't have any embedded
information in them. SUBNET would generate the
API key on cluster registration, and then it would
be saved in this config, to be used for subsequent
communication with SUBNET.
Harshavardhana
Poorna
Klaus Post
Krishnan Parthasarathi
Krishna Srinivas
Anis Elleuch
Andreas Auernhammer
Lenin Alevski
Xuehan Xu
hellivan
Shireesh Anjal
Daniel
Sidhartha Mani
Aditya Manthramurthy
yfanswer
Jarbitz
yinhen
Minio Trusted
jiangfucheng