zrepl/cmd/replication.go

320 lines
9.2 KiB
Go
Raw Permalink Normal View History

package cmd
import (
"fmt"
"io"
"time"
2017-09-22 14:13:58 +02:00
"bytes"
"encoding/json"
"github.com/zrepl/zrepl/rpc"
"github.com/zrepl/zrepl/util"
"github.com/zrepl/zrepl/zfs"
)
type localPullACL struct{}
func (a localPullACL) Filter(p *zfs.DatasetPath) (pass bool, err error) {
return true, nil
}
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
const LOCAL_TRANSPORT_IDENTITY string = "local"
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
const DEFAULT_INITIAL_REPL_POLICY = InitialReplPolicyMostRecent
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
type InitialReplPolicy string
const (
InitialReplPolicyMostRecent InitialReplPolicy = "most_recent"
InitialReplPolicyAll InitialReplPolicy = "all"
)
func closeRPCWithTimeout(log Logger, remote rpc.RPCClient, timeout time.Duration, goodbye string) {
log.Info("closing rpc connection")
ch := make(chan error)
go func() {
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
ch <- remote.Close()
close(ch)
}()
var err error
select {
case <-time.After(timeout):
err = fmt.Errorf("timeout exceeded (%s)", timeout)
case closeRequestErr := <-ch:
err = closeRequestErr
}
if err != nil {
log.WithError(err).Error("error closing connection")
}
return
}
type PullContext struct {
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
Remote rpc.RPCClient
Log Logger
Mapping DatasetMapping
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
InitialReplPolicy InitialReplPolicy
}
func doPull(pull PullContext) (err error) {
remote := pull.Remote
log := pull.Log
2017-09-22 14:13:58 +02:00
log.Info("request remote filesystem list")
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
fsr := FilesystemRequest{}
var remoteFilesystems []*zfs.DatasetPath
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
if err = remote.Call("FilesystemRequest", &fsr, &remoteFilesystems); err != nil {
return
}
2017-09-22 14:13:58 +02:00
log.Debug("map remote filesystems to local paths and determine order for per-filesystem sync")
type RemoteLocalMapping struct {
Remote *zfs.DatasetPath
Local *zfs.DatasetPath
}
replMapping := make(map[string]RemoteLocalMapping, len(remoteFilesystems))
localTraversal := zfs.NewDatasetPathForest()
for fs := range remoteFilesystems {
var err error
var localFs *zfs.DatasetPath
localFs, err = pull.Mapping.Map(remoteFilesystems[fs])
if err != nil {
2017-09-02 12:40:22 +02:00
err := fmt.Errorf("error mapping %s: %s", remoteFilesystems[fs], err)
log.WithError(err).WithField(logMapFromField, remoteFilesystems[fs]).Error("cannot map")
2017-09-02 12:40:22 +02:00
return err
}
if localFs == nil {
continue
}
log.WithField(logMapFromField, remoteFilesystems[fs].ToString()).
WithField(logMapToField, localFs.ToString()).Debug("mapping")
m := RemoteLocalMapping{remoteFilesystems[fs], localFs}
replMapping[m.Local.ToString()] = m
localTraversal.Add(m.Local)
}
2017-09-22 14:13:58 +02:00
log.Debug("build cache for already present local filesystem state")
localFilesystemState, err := zfs.ZFSListFilesystemState()
if err != nil {
2017-09-22 14:13:58 +02:00
log.WithError(err).Error("cannot request local filesystem state")
return err
}
2017-09-22 14:13:58 +02:00
log.Info("start per-filesystem sync")
localTraversal.WalkTopDown(func(v zfs.DatasetPathVisit) bool {
log := log.WithField(logFSField, v.Path.ToString())
2017-09-22 14:13:58 +02:00
if v.FilledIn {
if _, exists := localFilesystemState[v.Path.ToString()]; exists {
// No need to verify if this is a placeholder or not. It is sufficient
// to know we can add child filesystems to it
return true
}
2017-09-22 14:13:58 +02:00
log.Debug("create placeholder filesystem")
err = zfs.ZFSCreatePlaceholderFilesystem(v.Path)
if err != nil {
2017-09-22 14:13:58 +02:00
log.Error("cannot create placeholder filesystem")
return false
}
return true
}
m, ok := replMapping[v.Path.ToString()]
if !ok {
panic("internal inconsistency: replMapping should contain mapping for any path that was not filled in by WalkTopDown()")
}
log = log.WithField(logMapToField, m.Remote.ToString()).
WithField(logMapFromField, m.Local.ToString())
2017-09-22 14:13:58 +02:00
log.Debug("examing local filesystem state")
localState, localExists := localFilesystemState[m.Local.ToString()]
var versions []zfs.FilesystemVersion
switch {
case !localExists:
2017-09-22 14:13:58 +02:00
log.Info("local filesystem does not exist")
case localState.Placeholder:
2017-09-22 14:13:58 +02:00
log.Info("local filesystem is marked as placeholder")
default:
2017-09-22 14:13:58 +02:00
log.Debug("local filesystem exists")
log.Debug("requesting local filesystem versions")
if versions, err = zfs.ZFSListFilesystemVersions(m.Local, nil); err != nil {
2017-09-22 14:13:58 +02:00
log.WithError(err).Error("cannot get local filesystem versions")
return false
}
}
2017-09-22 14:13:58 +02:00
log.Info("requesting remote filesystem versions")
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
r := FilesystemVersionsRequest{
Filesystem: m.Remote,
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
}
var theirVersions []zfs.FilesystemVersion
if err = remote.Call("FilesystemVersionsRequest", &r, &theirVersions); err != nil {
2017-09-22 14:13:58 +02:00
log.WithError(err).Error("cannot get remote filesystem versions")
log.Warn("stopping replication for all filesystems mapped as children of receiving filesystem")
return false
}
2017-09-22 14:13:58 +02:00
log.Debug("computing diff between remote and local filesystem versions")
diff := zfs.MakeFilesystemDiff(versions, theirVersions)
2017-09-22 14:13:58 +02:00
log.WithField("diff", diff).Debug("diff between local and remote filesystem")
if localState.Placeholder && diff.Conflict != zfs.ConflictAllRight {
panic("internal inconsistency: local placeholder implies ConflictAllRight")
}
switch diff.Conflict {
case zfs.ConflictAllRight:
2017-09-22 14:13:58 +02:00
log.WithField("replication_policy", pull.InitialReplPolicy).Info("performing initial sync, following policy")
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
if pull.InitialReplPolicy != InitialReplPolicyMostRecent {
panic(fmt.Sprintf("policy '%s' not implemented", pull.InitialReplPolicy))
}
snapsOnly := make([]zfs.FilesystemVersion, 0, len(diff.MRCAPathRight))
for s := range diff.MRCAPathRight {
if diff.MRCAPathRight[s].Type == zfs.Snapshot {
snapsOnly = append(snapsOnly, diff.MRCAPathRight[s])
}
}
if len(snapsOnly) < 1 {
2017-09-22 14:13:58 +02:00
log.Warn("cannot perform initial sync: no remote snapshots")
return false
}
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
r := InitialTransferRequest{
Filesystem: m.Remote,
FilesystemVersion: snapsOnly[len(snapsOnly)-1],
}
log.WithField("version", r.FilesystemVersion).Debug("requesting snapshot stream")
var stream io.Reader
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
if err = remote.Call("InitialTransferRequest", &r, &stream); err != nil {
2017-09-22 14:13:58 +02:00
log.WithError(err).Error("cannot request initial transfer")
return false
}
2017-09-22 14:13:58 +02:00
log.Debug("received initial transfer request response")
2017-09-22 14:13:58 +02:00
log.Debug("invoke zfs receive")
watcher := util.IOProgressWatcher{Reader: stream}
watcher.KickOff(1*time.Second, func(p util.IOProgress) {
2017-09-22 14:13:58 +02:00
log.WithField("total_rx", p.TotalRX).Info("progress on receive operation")
})
recvArgs := []string{"-u"}
if localState.Placeholder {
2017-09-22 14:13:58 +02:00
log.Info("receive with forced rollback to replace placeholder filesystem")
recvArgs = append(recvArgs, "-F")
}
if err = zfs.ZFSRecv(m.Local, &watcher, recvArgs...); err != nil {
2017-09-22 14:13:58 +02:00
log.WithError(err).Error("canot receive stream")
return false
}
2017-09-22 14:13:58 +02:00
log.WithField("total_rx", watcher.Progress().TotalRX).
Info("finished receiving stream")
2017-09-22 14:13:58 +02:00
log.Debug("configuring properties of received filesystem")
if err = zfs.ZFSSet(m.Local, "readonly", "on"); err != nil {
2017-09-22 14:13:58 +02:00
log.WithError(err).Error("cannot set readonly property")
}
2017-09-22 14:13:58 +02:00
log.Info("finished initial transfer")
return true
case zfs.ConflictIncremental:
if len(diff.IncrementalPath) < 2 {
2017-09-22 14:13:58 +02:00
log.Info("remote and local are in sync")
return true
}
2017-09-22 14:13:58 +02:00
log.Info("following incremental path from diff")
var pathRx uint64
for i := 0; i < len(diff.IncrementalPath)-1; i++ {
from, to := diff.IncrementalPath[i], diff.IncrementalPath[i+1]
log, _ := log.WithField(logIncFromField, from.Name).WithField(logIncToField, to.Name), 0
2017-09-22 14:13:58 +02:00
log.Debug("requesting incremental snapshot stream")
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
r := IncrementalTransferRequest{
Filesystem: m.Remote,
From: from,
To: to,
}
var stream io.Reader
reimplement io.ReadWriteCloser based RPC mechanism The existing ByteStreamRPC requires writing RPC stub + server code for each RPC endpoint. Does not scale well. Goal: adding a new RPC call should - not require writing an RPC stub / handler - not require modifications to the RPC lib The wire format is inspired by HTTP2, the API by net/rpc. Frames are used for framing messages, i.e. a message is made of multiple frames which are glued together using a frame-bridging reader / writer. This roughly corresponds to HTTP2 streams, although we're happy with just one stream at any time and the resulting non-need for flow control, etc. Frames are typed using a header. The two most important types are 'Header' and 'Data'. The RPC protocol is built on top of this: - Client sends a header => multiple frames of type 'header' - Client sends request body => mulitiple frames of type 'data' - Server reads a header => multiple frames of type 'header' - Server reads request body => mulitiple frames of type 'data' - Server sends response header => ... - Server sends response body => ... An RPC header is serialized JSON and always the same structure. The body is of the type specified in the header. The RPC server and client use some semi-fancy reflection tequniques to automatically infer the data type of the request/response body based on the method signature of the server handler; or the client parameters, respectively. This boils down to a special-case for io.Reader, which are just dumped into a series of data frames as efficiently as possible. All other types are (de)serialized using encoding/json. The RPC layer and Frame Layer log some arbitrary messages that proved useful during debugging. By default, they log to a non-logger, which should not have a big impact on performance. pprof analysis shows the implementation spends its CPU time 60% waiting for syscalls 30% in memmove 10% ... On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the implementation achieved ~3.6GiB/s. Future optimization may include spice(2) / vmspice(2) on Linux, although this doesn't fit so well with the heavy use of io.Reader / io.Writer throughout the codebase. The existing hackaround for local calls was re-implemented to fit the new interface of PRCServer and RPCClient. The 'R'PC method invocation is a bit slower because reflection is involved inbetween, but otherwise performance should be no different. The RPC code currently does not support multipart requests and thus does not support the equivalent of a POST. Thus, the switch to the new rpc code had the following fallout: - Move request objects + constants from rpc package to main app code - Sacrifice the hacky 'push = pull me' way of doing push -> need to further extend RPC to support multipart requests or something to implement this properly with additional interfaces -> should be done after replication is abstracted better than separate algorithms for doPull() and doPush()
2017-08-19 22:37:14 +02:00
if err = remote.Call("IncrementalTransferRequest", &r, &stream); err != nil {
2017-09-22 14:13:58 +02:00
log.WithError(err).Error("cannot request incremental snapshot stream")
return false
}
2017-09-22 14:13:58 +02:00
log.Debug("invoking zfs receive")
watcher := util.IOProgressWatcher{Reader: stream}
watcher.KickOff(1*time.Second, func(p util.IOProgress) {
2017-09-22 14:13:58 +02:00
log.WithField("total_rx", p.TotalRX).Info("progress on receive operation")
})
if err = zfs.ZFSRecv(m.Local, &watcher); err != nil {
2017-09-22 14:13:58 +02:00
log.WithError(err).Error("cannot receive stream")
return false
}
totalRx := watcher.Progress().TotalRX
pathRx += totalRx
2017-09-22 14:13:58 +02:00
log.WithField("total_rx", totalRx).Info("finished incremental transfer")
}
2017-09-22 14:13:58 +02:00
log.WithField("total_rx", pathRx).Info("finished following incremental path")
return true
case zfs.ConflictNoCommonAncestor:
2017-09-22 14:13:58 +02:00
fallthrough
case zfs.ConflictDiverged:
2017-09-22 14:13:58 +02:00
var jsonDiff bytes.Buffer
if err := json.NewEncoder(&jsonDiff).Encode(diff); err != nil {
log.WithError(err).Error("cannot JSON-encode diff")
return false
}
2017-09-22 14:13:58 +02:00
var problem, resolution string
2017-09-22 14:13:58 +02:00
switch diff.Conflict {
case zfs.ConflictNoCommonAncestor:
problem = "remote and local filesystem have snapshots, but no common one"
resolution = "perform manual establish a common snapshot history"
case zfs.ConflictDiverged:
problem = "remote and local filesystem share a history but have diverged"
resolution = "perform manual replication or delete snapshots on the receiving" +
"side to establish an incremental replication parse"
}
2017-09-22 14:13:58 +02:00
log.WithField("diff", jsonDiff.String()).
WithField("problem", problem).
WithField("resolution", resolution).
Error("manual conflict resolution required")
return false
}
panic("should not be reached")
})
return
}