Replace the `io.Pipe` from streamingBitrotWriter -> CreateFile with a fixed size ring buffer.
This will add an output buffer for encoded shards to be written to disk - potentially via RPC.
This will remove blocking when `(*streamingBitrotWriter).Write` is called, and it writes hashes and data.
With current settings, the write looks like this:
```
Outbound
┌───────────────────┐ ┌────────────────┐ ┌───────────────┐ ┌────────────────┐
│ │ Parr. │ │ (http body) │ │ │ │
│ Bitrot Hash │ Write │ Pipe │ Read │ HTTP buffer │ Write (syscall) │ TCP Buffer │
│ Erasure Shard │ ──────────► │ (unbuffered) │ ────────────► │ (64K Max) │ ───────────────────► │ (4MB) │
│ │ │ │ │ (io.Copy) │ │ │
└───────────────────┘ └────────────────┘ └───────────────┘ └────────────────┘
```
We write a Hash (32 bytes). Since the pipe is unbuffered, it will block until the 32 bytes have
been delivered to the TCP buffer, and the next Read hits the Pipe.
Then we write the shard data. This will typically be bigger than 64KB, so it will block until two blocks
have been read from the pipe.
When we insert a ring buffer:
```
Outbound
┌───────────────────┐ ┌────────────────┐ ┌───────────────┐ ┌────────────────┐
│ │ │ │ (http body) │ │ │ │
│ Bitrot Hash │ Write │ Ring Buffer │ Read │ HTTP buffer │ Write (syscall) │ TCP Buffer │
│ Erasure Shard │ ──────────► │ (2MB) │ ────────────► │ (64K Max) │ ───────────────────► │ (4MB) │
│ │ │ │ │ (io.Copy) │ │ │
└───────────────────┘ └────────────────┘ └───────────────┘ └────────────────┘
```
The hash+shard will fit within the ring buffer, so writes will not block - but will complete after a
memcopy. Reads can fill the 64KB buffer if there is data for it.
If the network is congested, the ring buffer will become filled, and all syscalls will be on full buffers.
Only when the ring buffer is filled will erasure coding start blocking.
Since there is always "space" to write output data, we remove the parallel writing since we are
always writing to memory now, and the goroutine synchronization overhead probably not worth taking.
If the output were blocked in the existing, we would still wait for it to unblock in parallel write, so it would
make no difference there - except now the ring buffer smoothes out the load.
There are some micro-optimizations we could look at later. The biggest is that, in most cases,
we could encode directly to the ring buffer - if we are not at a boundary. Also, "force filling" the
Read requests (i.e., blocking until a full read can be completed) could be investigated and maybe
allow concurrent memory on read and write.
for actionable, inspections we have `mc support inspect`
we do not need double logging, healing will report relevant
errors if any, in terms of quorum lost etc.
Each Put, List, Multipart operations heavily rely on making
GetBucketInfo() call to verify if bucket exists or not on
a regular basis. This has a large performance cost when there
are tons of servers involved.
We did optimize this part by vectorizing the bucket calls,
however its not enough, beyond 100 nodes and this becomes
fairly visible in terms of performance.
#11878 added "keepHTTPResponseAlive" to CreateFile requests.
The problem is that it will begin writing to the response before the
body is read after 10 seconds. This will abort the writes on the
client-side, since it assumes the server has received what it wants.
The proposed solution here is to monitor the completion of the body
before beginning to send keepalive pings.
Fixes observed high number of goroutines stuck in `io.Copy` in
`github.com/minio/minio/cmd.(*xlStorage).CreateFile` and
`(*storageRESTClient).CreateFile` stuck in `http.DrainBody`.
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`
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.
- GetObject() should always use a common dataDir to
read from when it starts reading, this allows the
code in erasure decoding to have sane expectations.
- Healing should always heal on the common dataDir, this
allows the code in dangling object detection to purge
dangling content.
These both situations can happen under certain types of
retries during PUT when server is restarting etc, some
namespace entries might be left over.
wait groups are necessary with io.Pipes() to avoid
races when a blocking function may not be expected
and a Write() -> Close() before Read() races on each
other. We should avoid such situations..
Co-authored-by: Klaus Post <klauspost@gmail.com>
avoid time_wait build up with getObject requests if there are
pending callers and they timeout, can lead to time_wait states
Bonus share the same buffer pool with erasure healing logic,
additionally also fixes a race where parallel readers were
never cleanup during Encode() phase, because pipe.Reader end
was never closed().
Added closer right away upon an error during Encode to make
sure to avoid racy Close() while stream was still being
Read().
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.
Previously we added heal trigger when bit-rot checks
failed, now extend that to support heal when parts
are not found either. This healing gets only triggered
if we can successfully decode the object i.e read
quorum is still satisfied for the object.
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.
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.
- Heal if the part.1 is truncated from its original size
- Heal if the part.1 fails while being verified in between
- Heal if the part.1 fails while being at a certain offset
Other cleanups include make sure to flush the HTTP responses
properly from storage-rest-server, avoid using 'defer' to
improve call latency. 'defer' incurs latency avoid them
in our hot-paths such as storage-rest handlers.
Fixes#8319
