哨兵、主从、集群,一串下来。
redis cluster 主要作用如下(虽然是显而易见了):
数据分片,流量分发。
cluster 将不同的数据分发给不同的节点,不过没有使用一致性hash算法,而是引入了Hash槽位的概念。cluster有16384个槽位,每个槽位只能指派给一个节点。
所以本文的重心也就很明确了:redis集群是如何实现通信及数据分片、流量分发的
关于实操:【redis】闲得无聊,来聊聊当下爆火的 redis集群,顺便搭一个玩玩呗
如果cluster中的某个节点收到客户端请求,但请求中查询的键不是当前节点负责的,则它将通知客户端进行重定向,客户端重新发送请求给真正的数据存储节点。
那是怎么实现?包发过来,不在我这儿,我告诉她在他那儿,让她重发给他?
还是包发过来,不在我这儿,我告诉她不在我这儿,因为我也不知道在谁那儿,让她一个一个自己去试试?
还是什么其他的方式?
我想是第一种,但是现实是什么样的,再看看。
int processCommand(client *c) {
......
/* If cluster is enabled perform the cluster redirection here.
* However we don't perform the redirection if:
* 1) The sender of this command is our master.
* 2) The command has no key arguments. */
if (server.cluster_enabled &&
!(c->flags & CLIENT_MASTER) &&
!(c->flags & CLIENT_LUA &&
server.lua_caller->flags & CLIENT_MASTER) &&
!(c->cmd->getkeys_proc == NULL && c->cmd->firstkey == 0 &&
c->cmd->proc != execCommand))
{
int hashslot;
int error_code;
//查找真正的存储节点
clusterNode *n = getNodeByQuery(c,c->cmd,c->argv,c->argc,
&hashslot,&error_code);
if (n == NULL || n != server.cluster->myself) {
if (c->cmd->proc == execCommand) {
discardTransaction(c);
} else {
flagTransaction(c);
}
//返回 ASK 或 MOBED 转向标志及重定向目标节点,通知客户端重定向
/*
如果对应槽位数据正在迁出,则返回 ASK,提示客户端仅在下一条命令中重定向目标节点。
否则返回MOVED,提示客户端可以长期重定向。
*/
clusterRedirectClient(c,n,hashslot,error_code);
return C_OK;
}
}
......
}
getNodeByQuery函数负责查找数据存储节点:
/* Return the pointer to the cluster node that is able to serve the command.
* For the function to succeed the command should only target either:
*
* 1) A single key (even multiple times like LPOPRPUSH mylist mylist).
* 2) Multiple keys in the same hash slot, while the slot is stable (no
* resharding in progress).
*
* On success the function returns the node that is able to serve the request.
* If the node is not 'myself' a redirection must be perfomed. The kind of
* redirection is specified setting the integer passed by reference
* 'error_code', which will be set to CLUSTER_REDIR_ASK or
* CLUSTER_REDIR_MOVED.
*
* When the node is 'myself' 'error_code' is set to CLUSTER_REDIR_NONE.
*
* If the command fails NULL is returned, and the reason of the failure is
* provided via 'error_code', which will be set to:
*
* CLUSTER_REDIR_CROSS_SLOT if the request contains multiple keys that
* don't belong to the same hash slot.
*
* CLUSTER_REDIR_UNSTABLE if the request contains multiple keys
* belonging to the same slot, but the slot is not stable (in migration or
* importing state, likely because a resharding is in progress).
*
* CLUSTER_REDIR_DOWN_UNBOUND if the request addresses a slot which is
* not bound to any node. In this case the cluster global state should be
* already "down" but it is fragile to rely on the update of the global state,
* so we also handle it here.
*
* CLUSTER_REDIR_DOWN_STATE and CLUSTER_REDIR_DOWN_RO_STATE if the cluster is
* down but the user attempts to execute a command that addresses one or more keys. */
clusterNode *getNodeByQuery(client *c, struct redisCommand *cmd, robj **argv, int argc, int *hashslot, int *error_code) {
clusterNode *n = NULL;
robj *firstkey = NULL;
int multiple_keys = 0;
multiState *ms, _ms;
multiCmd mc;
int i, slot = 0, migrating_slot = 0, importing_slot = 0, missing_keys = 0;
/* Allow any key to be set if a module disabled cluster redirections. */
if (server.cluster_module_flags & CLUSTER_MODULE_FLAG_NO_REDIRECTION)
return myself;
/* Set error code optimistically for the base case. */
if (error_code) *error_code = CLUSTER_REDIR_NONE;
/* Modules can turn off Redis Cluster redirection: this is useful
* when writing a module that implements a completely different
* distributed system. */
/* We handle all the cases as if they were EXEC commands, so we have
* a common code path for everything */
if (cmd->proc == execCommand) {
/* If CLIENT_MULTI flag is not set EXEC is just going to return an
* error. */
if (!(c->flags & CLIENT_MULTI)) return myself;
ms = &c->mstate;
} else {
/* In order to have a single codepath create a fake Multi State
* structure if the client is not in MULTI/EXEC state, this way
* we have a single codepath below. */
ms = &_ms;
_ms.commands = &mc;
_ms.count = 1;
mc.argv = argv;
mc.argc = argc;
mc.cmd = cmd;
}
/* Check that all the keys are in the same hash slot, and obtain this
* slot and the node associated. */
for (i = 0; i < ms->count; i++) {
struct redisCommand *mcmd;
robj **margv;
int margc, *keyindex, numkeys, j;
mcmd = ms->commands[i].cmd;
margc = ms->commands[i].argc;
margv = ms->commands[i].argv;
keyindex = getKeysFromCommand(mcmd,margv,margc,&numkeys);
for (j = 0; j < numkeys; j++) {
robj *thiskey = margv[keyindex[j]];
int thisslot = keyHashSlot((char*)thiskey->ptr,
sdslen(thiskey->ptr));
if (firstkey == NULL) {
/* This is the first key we see. Check what is the slot
* and node. */
firstkey = thiskey;
slot = thisslot;
n = server.cluster->slots[slot];
/* Error: If a slot is not served, we are in "cluster down"
* state. However the state is yet to be updated, so this was
* not trapped earlier in processCommand(). Report the same
* error to the client. */
if (n == NULL) {
getKeysFreeResult(keyindex);
if (error_code)
*error_code = CLUSTER_REDIR_DOWN_UNBOUND;
return NULL;
}
/* If we are migrating or importing this slot, we need to check
* if we have all the keys in the request (the only way we
* can safely serve the request, otherwise we return a TRYAGAIN
* error). To do so we set the importing/migrating state and
* increment a counter for every missing key. */
if (n == myself &&
server.cluster->migrating_slots_to[slot] != NULL)
{
migrating_slot = 1;
} else if (server.cluster->importing_slots_from[slot] != NULL) {
importing_slot = 1;
}
} else {
/* If it is not the first key, make sure it is exactly
* the same key as the first we saw. */
if (!equalStringObjects(firstkey,thiskey)) {
if (slot != thisslot) {
/* Error: multiple keys from different slots. */
getKeysFreeResult(keyindex);
if (error_code)
*error_code = CLUSTER_REDIR_CROSS_SLOT;
return NULL;
} else {
/* Flag this request as one with multiple different
* keys. */
multiple_keys = 1;
}
}
}
/* Migarting / Improrting slot? Count keys we don't have. */
if ((migrating_slot || importing_slot) &&
lookupKeyRead(&server.db[0],thiskey) == NULL)
{
missing_keys++;
}
}
getKeysFreeResult(keyindex);
}
/* No key at all in command? then we can serve the request
* without redirections or errors in all the cases. */
if (n == NULL) return myself;
/* Cluster is globally down but we got keys? We only serve the request
* if it is a read command and when allow_reads_when_down is enabled. */
if (server.cluster->state != CLUSTER_OK) {
if (!server.cluster_allow_reads_when_down) {
/* The cluster is configured to block commands when the
* cluster is down. */
if (error_code) *error_code = CLUSTER_REDIR_DOWN_STATE;
return NULL;
} else if (!(cmd->flags & CMD_READONLY) && !(cmd->proc == evalCommand)
&& !(cmd->proc == evalShaCommand))
{
/* The cluster is configured to allow read only commands
* but this command is neither readonly, nor EVAL or
* EVALSHA. */
if (error_code) *error_code = CLUSTER_REDIR_DOWN_RO_STATE;
return NULL;
} else {
/* Fall through and allow the command to be executed:
* this happens when server.cluster_allow_reads_when_down is
* true and the command is a readonly command or EVAL / EVALSHA. */
}
}
/* Return the hashslot by reference. */
if (hashslot) *hashslot = slot;
/* MIGRATE always works in the context of the local node if the slot
* is open (migrating or importing state). We need to be able to freely
* move keys among instances in this case. */
if ((migrating_slot || importing_slot) && cmd->proc == migrateCommand)
return myself;
/* If we don't have all the keys and we are migrating the slot, send
* an ASK redirection. */
if (migrating_slot && missing_keys) {
if (error_code) *error_code = CLUSTER_REDIR_ASK;
return server.cluster->migrating_slots_to[slot];
}
/* If we are receiving the slot, and the client correctly flagged the
* request as "ASKING", we can serve the request. However if the request
* involves multiple keys and we don't have them all, the only option is
* to send a TRYAGAIN error. */
if (importing_slot &&
(c->flags & CLIENT_ASKING || cmd->flags & CMD_ASKING))
{
if (multiple_keys && missing_keys) {
if (error_code) *error_code = CLUSTER_REDIR_UNSTABLE;
return NULL;
} else {
return myself;
}
}
/* Handle the read-only client case reading from a slave: if this
* node is a slave and the request is about an hash slot our master
* is serving, we can reply without redirection. */
if (c->flags & CLIENT_READONLY &&
(cmd->flags & CMD_READONLY || cmd->proc == evalCommand ||
cmd->proc == evalShaCommand) &&
nodeIsSlave(myself) &&
myself->slaveof == n)
{
return myself;
}
/* Base case: just return the right node. However if this node is not
* myself, set error_code to MOVED since we need to issue a rediretion. */
if (n != myself && error_code) *error_code = CLUSTER_REDIR_MOVED;
return n;
}
什么是槽位迁移?rehash,就这样理解嘛。
什么时候发生?比方我某个节点挂了、比方我某个节点又复活可以分摊压力了;
/* MIGRATE host port key dbid timeout [COPY | REPLACE | AUTH password |
* AUTH2 username password]
*
* On in the multiple keys form:
*
* MIGRATE host port "" dbid timeout [COPY | REPLACE | AUTH password |
* AUTH2 username password] KEYS key1 key2 ... keyN */
void migrateCommand(client *c) {
migrateCachedSocket *cs;
int copy = 0, replace = 0, j;
char *username = NULL;
char *password = NULL;
long timeout;
long dbid;
robj **ov = NULL; /* Objects to migrate. */
robj **kv = NULL; /* Key names. */
robj **newargv = NULL; /* Used to rewrite the command as DEL ... keys ... */
rio cmd, payload;
int may_retry = 1;
int write_error = 0;
int argv_rewritten = 0;
/* To support the KEYS option we need the following additional state. */
int first_key = 3; /* Argument index of the first key. */
int num_keys = 1; /* By default only migrate the 'key' argument. */
/* Parse additional options */
for (j = 6; j < c->argc; j++) {
int moreargs = (c->argc-1) - j;
if (!strcasecmp(c->argv[j]->ptr,"copy")) {
copy = 1;
} else if (!strcasecmp(c->argv[j]->ptr,"replace")) {
replace = 1;
} else if (!strcasecmp(c->argv[j]->ptr,"auth")) {
if (!moreargs) {
addReply(c,shared.syntaxerr);
return;
}
j++;
password = c->argv[j]->ptr;
} else if (!strcasecmp(c->argv[j]->ptr,"auth2")) {
if (moreargs < 2) {
addReply(c,shared.syntaxerr);
return;
}
username = c->argv[++j]->ptr;
password = c->argv[++j]->ptr;
} else if (!strcasecmp(c->argv[j]->ptr,"keys")) {
if (sdslen(c->argv[3]->ptr) != 0) {
addReplyError(c,
"When using MIGRATE KEYS option, the key argument"
" must be set to the empty string");
return;
}
first_key = j+1;
num_keys = c->argc - j - 1;
break; /* All the remaining args are keys. */
} else {
addReply(c,shared.syntaxerr);
return;
}
}
/* Sanity check */
if (getLongFromObjectOrReply(c,c->argv[5],&timeout,NULL) != C_OK ||
getLongFromObjectOrReply(c,c->argv[4],&dbid,NULL) != C_OK)
{
return;
}
if (timeout <= 0) timeout = 1000;
/* Check if the keys are here. If at least one key is to migrate, do it
* otherwise if all the keys are missing reply with "NOKEY" to signal
* the caller there was nothing to migrate. We don't return an error in
* this case, since often this is due to a normal condition like the key
* expiring in the meantime. */
ov = zrealloc(ov,sizeof(robj*)*num_keys);
kv = zrealloc(kv,sizeof(robj*)*num_keys);
int oi = 0;
for (j = 0; j < num_keys; j++) {
if ((ov[oi] = lookupKeyRead(c->db,c->argv[first_key+j])) != NULL) {
kv[oi] = c->argv[first_key+j];
oi++;
}
}
num_keys = oi;
if (num_keys == 0) {
zfree(ov); zfree(kv);
addReplySds(c,sdsnew("+NOKEY\r\n"));
return;
}
try_again:
write_error = 0;
/* Connect */
cs = migrateGetSocket(c,c->argv[1],c->argv[2],timeout);
if (cs == NULL) {
zfree(ov); zfree(kv);
return; /* error sent to the client by migrateGetSocket() */
}
rioInitWithBuffer(&cmd,sdsempty());
/* Authentication */
if (password) {
int arity = username ? 3 : 2;
serverAssertWithInfo(c,NULL,rioWriteBulkCount(&cmd,'*',arity));
serverAssertWithInfo(c,NULL,rioWriteBulkString(&cmd,"AUTH",4));
if (username) {
serverAssertWithInfo(c,NULL,rioWriteBulkString(&cmd,username,
sdslen(username)));
}
serverAssertWithInfo(c,NULL,rioWriteBulkString(&cmd,password,
sdslen(password)));
}
/* Send the SELECT command if the current DB is not already selected. */
int select = cs->last_dbid != dbid; /* Should we emit SELECT? */
if (select) {
serverAssertWithInfo(c,NULL,rioWriteBulkCount(&cmd,'*',2));
serverAssertWithInfo(c,NULL,rioWriteBulkString(&cmd,"SELECT",6));
serverAssertWithInfo(c,NULL,rioWriteBulkLongLong(&cmd,dbid));
}
int non_expired = 0; /* Number of keys that we'll find non expired.
Note that serializing large keys may take some time
so certain keys that were found non expired by the
lookupKey() function, may be expired later. */
/* Create RESTORE payload and generate the protocol to call the command. */
for (j = 0; j < num_keys; j++) {
long long ttl = 0;
long long expireat = getExpire(c->db,kv[j]);
if (expireat != -1) {
ttl = expireat-mstime();
if (ttl < 0) {
continue;
}
if (ttl < 1) ttl = 1;
}
/* Relocate valid (non expired) keys into the array in successive
* positions to remove holes created by the keys that were present
* in the first lookup but are now expired after the second lookup. */
kv[non_expired++] = kv[j];
serverAssertWithInfo(c,NULL,
rioWriteBulkCount(&cmd,'*',replace ? 5 : 4));
if (server.cluster_enabled)
serverAssertWithInfo(c,NULL,
rioWriteBulkString(&cmd,"RESTORE-ASKING",14));
else
serverAssertWithInfo(c,NULL,rioWriteBulkString(&cmd,"RESTORE",7));
serverAssertWithInfo(c,NULL,sdsEncodedObject(kv[j]));
serverAssertWithInfo(c,NULL,rioWriteBulkString(&cmd,kv[j]->ptr,
sdslen(kv[j]->ptr)));
serverAssertWithInfo(c,NULL,rioWriteBulkLongLong(&cmd,ttl));
/* Emit the payload argument, that is the serialized object using
* the DUMP format. */
createDumpPayload(&payload,ov[j],kv[j]);
serverAssertWithInfo(c,NULL,
rioWriteBulkString(&cmd,payload.io.buffer.ptr,
sdslen(payload.io.buffer.ptr)));
sdsfree(payload.io.buffer.ptr);
/* Add the REPLACE option to the RESTORE command if it was specified
* as a MIGRATE option. */
if (replace)
serverAssertWithInfo(c,NULL,rioWriteBulkString(&cmd,"REPLACE",7));
}
/* Fix the actual number of keys we are migrating. */
num_keys = non_expired;
/* Transfer the query to the other node in 64K chunks. */
errno = 0;
{
sds buf = cmd.io.buffer.ptr;
size_t pos = 0, towrite;
int nwritten = 0;
while ((towrite = sdslen(buf)-pos) > 0) {
towrite = (towrite > (64*1024) ? (64*1024) : towrite);
nwritten = connSyncWrite(cs->conn,buf+pos,towrite,timeout);
if (nwritten != (signed)towrite) {
write_error = 1;
goto socket_err;
}
pos += nwritten;
}
}
char buf0[1024]; /* Auth reply. */
char buf1[1024]; /* Select reply. */
char buf2[1024]; /* Restore reply. */
/* Read the AUTH reply if needed. */
if (password && connSyncReadLine(cs->conn, buf0, sizeof(buf0), timeout) <= 0)
goto socket_err;
/* Read the SELECT reply if needed. */
if (select && connSyncReadLine(cs->conn, buf1, sizeof(buf1), timeout) <= 0)
goto socket_err;
/* Read the RESTORE replies. */
int error_from_target = 0;
int socket_error = 0;
int del_idx = 1; /* Index of the key argument for the replicated DEL op. */
/* Allocate the new argument vector that will replace the current command,
* to propagate the MIGRATE as a DEL command (if no COPY option was given).
* We allocate num_keys+1 because the additional argument is for "DEL"
* command name itself. */
if (!copy) newargv = zmalloc(sizeof(robj*)*(num_keys+1));
for (j = 0; j < num_keys; j++) {
if (connSyncReadLine(cs->conn, buf2, sizeof(buf2), timeout) <= 0) {
socket_error = 1;
break;
}
if ((password && buf0[0] == '-') ||
(select && buf1[0] == '-') ||
buf2[0] == '-')
{
/* On error assume that last_dbid is no longer valid. */
if (!error_from_target) {
cs->last_dbid = -1;
char *errbuf;
if (password && buf0[0] == '-') errbuf = buf0;
else if (select && buf1[0] == '-') errbuf = buf1;
else errbuf = buf2;
error_from_target = 1;
addReplyErrorFormat(c,"Target instance replied with error: %s",
errbuf+1);
}
} else {
if (!copy) {
/* No COPY option: remove the local key, signal the change. */
dbDelete(c->db,kv[j]);
signalModifiedKey(c,c->db,kv[j]);
notifyKeyspaceEvent(NOTIFY_GENERIC,"del",kv[j],c->db->id);
server.dirty++;
/* Populate the argument vector to replace the old one. */
newargv[del_idx++] = kv[j];
incrRefCount(kv[j]);
}
}
}
/* On socket error, if we want to retry, do it now before rewriting the
* command vector. We only retry if we are sure nothing was processed
* and we failed to read the first reply (j == 0 test). */
if (!error_from_target && socket_error && j == 0 && may_retry &&
errno != ETIMEDOUT)
{
goto socket_err; /* A retry is guaranteed because of tested conditions.*/
}
/* On socket errors, close the migration socket now that we still have
* the original host/port in the ARGV. Later the original command may be
* rewritten to DEL and will be too later. */
if (socket_error) migrateCloseSocket(c->argv[1],c->argv[2]);
if (!copy) {
/* Translate MIGRATE as DEL for replication/AOF. Note that we do
* this only for the keys for which we received an acknowledgement
* from the receiving Redis server, by using the del_idx index. */
if (del_idx > 1) {
newargv[0] = createStringObject("DEL",3);
/* Note that the following call takes ownership of newargv. */
replaceClientCommandVector(c,del_idx,newargv);
argv_rewritten = 1;
} else {
/* No key transfer acknowledged, no need to rewrite as DEL. */
zfree(newargv);
}
newargv = NULL; /* Make it safe to call zfree() on it in the future. */
}
/* If we are here and a socket error happened, we don't want to retry.
* Just signal the problem to the client, but only do it if we did not
* already queue a different error reported by the destination server. */
if (!error_from_target && socket_error) {
may_retry = 0;
goto socket_err;
}
if (!error_from_target) {
/* Success! Update the last_dbid in migrateCachedSocket, so that we can
* avoid SELECT the next time if the target DB is the same. Reply +OK.
*
* Note: If we reached this point, even if socket_error is true
* still the SELECT command succeeded (otherwise the code jumps to
* socket_err label. */
cs->last_dbid = dbid;
addReply(c,shared.ok);
} else {
/* On error we already sent it in the for loop above, and set
* the currently selected socket to -1 to force SELECT the next time. */
}
sdsfree(cmd.io.buffer.ptr);
zfree(ov); zfree(kv); zfree(newargv);
return;
/* On socket errors we try to close the cached socket and try again.
* It is very common for the cached socket to get closed, if just reopening
* it works it's a shame to notify the error to the caller. */
socket_err:
/* Cleanup we want to perform in both the retry and no retry case.
* Note: Closing the migrate socket will also force SELECT next time. */
sdsfree(cmd.io.buffer.ptr);
/* If the command was rewritten as DEL and there was a socket error,
* we already closed the socket earlier. While migrateCloseSocket()
* is idempotent, the host/port arguments are now gone, so don't do it
* again. */
if (!argv_rewritten) migrateCloseSocket(c->argv[1],c->argv[2]);
zfree(newargv);
newargv = NULL; /* This will get reallocated on retry. */
/* Retry only if it's not a timeout and we never attempted a retry
* (or the code jumping here did not set may_retry to zero). */
if (errno != ETIMEDOUT && may_retry) {
may_retry = 0;
goto try_again;
}
/* Cleanup we want to do if no retry is attempted. */
zfree(ov); zfree(kv);
addReplySds(c,
sdscatprintf(sdsempty(),
"-IOERR error or timeout %s to target instance\r\n",
write_error ? "writing" : "reading"));
return;
}
这个启动过程我就不放代码了,和前面的主从、哨兵有很多相似之处。
1、节点启动
2、节点握手,参见主从握手
3、指派槽位
4、建立主从关系
5、节点通信:Gossip算法
6、故障转移,故障转移机制与哨兵基本一致