【Redis】RDB持久化

RDB是redis中的一种持久化方式,以二进制形式存储在文件中,且排列非常紧凑,这也意味着文件更小,我们可以更快的载入数据,但其也有坏处,就是及其容易丢失数据,.因为其会遵循配置文件中默认的配置X秒Y条数据的写入时会执行RDB,但未满足时便一直存在内存中,如果此时服务器宕机甚至断电,距上一次持久化到现在的数据都将丢失.因为其在持久化时会把所有的数据全部写入一个临时文件,然后改名,在后台执行时还会fork,这使得在有大量数据的时候很慢.但它也有好处,就是可以更快的恢复,这使得它在做副本的时候是一个很好的选择.

通过SAVE或者BGSAVE操作可以显式调用RDB.SAVE创建RDB文件的速度会比BGSAVE快,SAVE可以集中资源来创建RDB文件,如果数据库正在上线当中.就要使用BGSAVE,如果数据库需要维护,可以使用SAVE命令.

RDB(后台)基本的执行过程是这样的

1. 子进程把数据写入临时文件 结束的话向父进程发送信号

2. 父进程接收到信号以后更新属性 然后改名 覆盖原来的dump.rdb

系统初始化后会周期执行 serverCron() 函数，其中会调用 rdb 持久化的入口函数  rdbSaveBackground()。

rdb.c rdbSaveBackground() 函数如下：
int rdbSaveBackground(char *filename, rdbSaveInfo *rsi) {
    pid_t childpid;
    if (hasActiveChildProcess()) return C_ERR;
    server.dirty_before_bgsave = server.dirty;
    server.lastbgsave_try = time(NULL);
    openChildInfoPipe();
    // 创建个子进程来执行 rdbSave() 生成临时文件
    if ((childpid = redisFork(CHILD_TYPE_RDB)) == 0) {
        int retval;
        /* Child */
        redisSetProcTitle("redis-rdb-bgsave");
        redisSetCpuAffinity(server.bgsave_cpulist);
        retval = rdbSave(filename,rsi);
        if (retval == C_OK) {
            sendChildCOWInfo(CHILD_TYPE_RDB, "RDB");
        }
        // 子进程执行临时文件生成后推出来向父线程发信号
        exitFromChild((retval == C_OK) ? 0 : 1);
    } else {
        /* Parent */
        if (childpid == -1) {
            closeChildInfoPipe();
            server.lastbgsave_status = C_ERR;
            serverLog(LL_WARNING,"Can't save in background: fork: %s",
                strerror(errno));
            return C_ERR;
        }
        serverLog(LL_NOTICE,"Background saving started by pid %d",childpid);
        server.rdb_save_time_start = time(NULL);
        server.rdb_child_pid = childpid;
        server.rdb_child_type = RDB_CHILD_TYPE_DISK;
        updateDictResizePolicy();
        return C_OK;
    }
    return C_OK; /* unreached */
}

子进程会开始 rdb 临时文件生成，主要调用 rdbSave() 函数：

rdb.c rdbSave() 并进一步调用 rdbSaveRio()，会根据 rdb 文件协议生成临时文件，遍历所有数据库：

rdb.c rdbSaveRio() 函数如下：

int rdbSaveRio(rio *rdb, int *error, int rdbflags, rdbSaveInfo *rsi) {
    dictIterator *di = NULL;
    dictEntry *de;
    char magic[10];
    int j;
    uint64_t cksum;
    size_t processed = 0;
    if (server.rdb_checksum)
        rdb->update_cksum = rioGenericUpdateChecksum;
    // 文件起始字符串
    snprintf(magic,sizeof(magic),"REDIS%04d",RDB_VERSION);
    if (rdbWriteRaw(rdb,magic,9) == -1) goto werr;
    if (rdbSaveInfoAuxFields(rdb,rdbflags,rsi) == -1) goto werr;
    if (rdbSaveModulesAux(rdb, REDISMODULE_AUX_BEFORE_RDB) == -1) goto werr;
    for (j = 0; j < server.dbnum; j++) {
        redisDb *db = server.db+j;
        dict *d = db->dict;
        if (dictSize(d) == 0) continue;
        di = dictGetSafeIterator(d);
        /* Write the SELECT DB opcode */
        if (rdbSaveType(rdb,RDB_OPCODE_SELECTDB) == -1) goto werr;
        if (rdbSaveLen(rdb,j) == -1) goto werr;
        /* Write the RESIZE DB opcode. */
        uint64_t db_size, expires_size;
        db_size = dictSize(db->dict);
        expires_size = dictSize(db->expires);
        if (rdbSaveType(rdb,RDB_OPCODE_RESIZEDB) == -1) goto werr;
        if (rdbSaveLen(rdb,db_size) == -1) goto werr;
        if (rdbSaveLen(rdb,expires_size) == -1) goto werr;
        /* Iterate this DB writing every entry */
        while((de = dictNext(di)) != NULL) {
            sds keystr = dictGetKey(de);
            robj key, *o = dictGetVal(de);
            long long expire;
            initStaticStringObject(key,keystr);
            expire = getExpire(db,&key);
            if (rdbSaveKeyValuePair(rdb,&key,o,expire) == -1) goto werr;
            /* When this RDB is produced as part of an AOF rewrite, move
             * accumulated diff from parent to child while rewriting in
             * order to have a smaller final write. */
            if (rdbflags & RDBFLAGS_AOF_PREAMBLE &&
                rdb->processed_bytes > processed+AOF_READ_DIFF_INTERVAL_BYTES)
            {
                processed = rdb->processed_bytes;
                aofReadDiffFromParent();
            }
        }
        dictReleaseIterator(di);
        di = NULL; /* So that we don't release it again on error. */
    }
    /* If we are storing the replication information on disk, persist
     * the script cache as well: on successful PSYNC after a restart, we need
     * to be able to process any EVALSHA inside the replication backlog the
     * master will send us. */
    if (rsi && dictSize(server.lua_scripts)) {
        di = dictGetIterator(server.lua_scripts);
        while((de = dictNext(di)) != NULL) {
            robj *body = dictGetVal(de);
            if (rdbSaveAuxField(rdb,"lua",3,body->ptr,sdslen(body->ptr)) == -1)
                goto werr;
        }
        dictReleaseIterator(di);
        di = NULL; /* So that we don't release it again on error. */
    }
    if (rdbSaveModulesAux(rdb, REDISMODULE_AUX_AFTER_RDB) == -1) goto werr;
    /* EOF opcode */
    if (rdbSaveType(rdb,RDB_OPCODE_EOF) == -1) goto werr;
    /* CRC64 checksum. It will be zero if checksum computation is disabled, the
     * loading code skips the check in this case. */
    cksum = rdb->cksum;
    memrev64ifbe(&cksum);
    if (rioWrite(rdb,&cksum,8) == 0) goto werr;
    return C_OK;
werr:
    if (error) *error = errno;
    if (di) dictReleaseIterator(di);
    return C_ERR;
}

具体是根据 rdb 协议将 ky 进行写入：

rdb.c rdbSaveKeyValuePair() 函数如下：

int rdbSaveKeyValuePair(rio *rdb, robj *key, robj *val, long long expiretime) {
    int savelru = server.maxmemory_policy & MAXMEMORY_FLAG_LRU;
    int savelfu = server.maxmemory_policy & MAXMEMORY_FLAG_LFU;
    /* Save the expire time */
    if (expiretime != -1) {
        if (rdbSaveType(rdb,RDB_OPCODE_EXPIRETIME_MS) == -1) return -1;
        if (rdbSaveMillisecondTime(rdb,expiretime) == -1) return -1;
    }
    /* Save the LRU info. */
    if (savelru) {
        uint64_t idletime = estimateObjectIdleTime(val);
        idletime /= 1000; /* Using seconds is enough and requires less space.*/
        if (rdbSaveType(rdb,RDB_OPCODE_IDLE) == -1) return -1;
        if (rdbSaveLen(rdb,idletime) == -1) return -1;
    }
    /* Save the LFU info. */
    if (savelfu) {
        uint8_t buf[1];
        buf[0] = LFUDecrAndReturn(val);
        /* We can encode this in exactly two bytes: the opcode and an 8
         * bit counter, since the frequency is logarithmic with a 0-255 range.
         * Note that we do not store the halving time because to reset it
         * a single time when loading does not affect the frequency much. */
        if (rdbSaveType(rdb,RDB_OPCODE_FREQ) == -1) return -1;
        if (rdbWriteRaw(rdb,buf,1) == -1) return -1;
    }
    /* Save type, key, value */
    if (rdbSaveObjectType(rdb,val) == -1) return -1;
    if (rdbSaveStringObject(rdb,key) == -1) return -1;
    if (rdbSaveObject(rdb,val,key) == -1) return -1;
    /* Delay return if required (for testing) */
    if (server.rdb_key_save_delay)
        usleep(server.rdb_key_save_delay);
    return 1;
}

父进程接受到信号之后将临时文件修改为 rdb 文件：

rdb.c backgroundSaveDoneHandler() 函数如下：

static void backgroundSaveDoneHandlerDisk(int exitcode, int bysignal) {
    if (!bysignal && exitcode == 0) {
        serverLog(LL_NOTICE,
            "Background saving terminated with success");
        server.dirty = server.dirty - server.dirty_before_bgsave;
        server.lastsave = time(NULL);
        server.lastbgsave_status = C_OK;
    } else if (!bysignal && exitcode != 0) {
        serverLog(LL_WARNING, "Background saving error");
        server.lastbgsave_status = C_ERR;
    } else {
        mstime_t latency;
        serverLog(LL_WARNING,
            "Background saving terminated by signal %d", bysignal);
        latencyStartMonitor(latency);
        rdbRemoveTempFile(server.rdb_child_pid, 0);
        latencyEndMonitor(latency);
        latencyAddSampleIfNeeded("rdb-unlink-temp-file",latency);
        /* SIGUSR1 is whitelisted, so we have a way to kill a child without
         * triggering an error condition. */
        if (bysignal != SIGUSR1)
            server.lastbgsave_status = C_ERR;
    }
}

整个运行流程如下：

1994:M 03 Feb 2023 17:14:36.961 * 1 changes in 5 seconds. Saving...

【log】rdb.c rdbSaveBackground() 后台rdb持久化任务

【log】server.c redisFork() 创建子进程

1994:M 03 Feb 2023 17:14:36.972 * Background saving started by pid 1995

【log】server.c redisSetCpuAffinity() 绑定进程和cpu,当前进程:42949672976,配置的cpu:(null)

【log】rdb.c rdbSave() 开始rdb持久化,filename=dump.rdb

【log】rdb.c rdbSaveRio() 根据ky和rdb协议生成临时文件

【log】rdb.c rdbSaveKeyValuePair() 将内存ky写入 rdb 文件

1995:C 03 Feb 2023 17:14:36.986 * DB saved on disk

【log】rdb.c backgroundSaveDoneHandler() 转换rdb临时文件为持久化文件

1994:M 03 Feb 2023 17:14:37.087 * Background saving terminated with success