Merge pull request #23465 from taosdata/fix/TD-26189-last-tier-write

enh(tsdb/retention): enable last tier write

source/common/src/tglobal.c

@@ -280,8 +280,10 @@ int8_t tsS3Enabled = false;
 int8_t tsS3Https = true;
 char   tsS3Hostname[TSDB_FQDN_LEN] = "<hostname>";
 
-int32_t tsS3BlockSize = 4096;     // number of tsdb pages
+int32_t tsS3BlockSize = -1;        // number of tsdb pages (4096)
 int32_t tsS3BlockCacheSize = 16;   // number of blocks
+int32_t tsS3PageCacheSize = 4096;  // number of pages
+int32_t tsS3UploadDelaySec = 60 * 60;
 
 #ifndef _STORAGE
 int32_t taosSetTfsCfg(SConfig *pCfg) {
@@ -460,7 +462,8 @@ static int32_t taosAddClientCfg(SConfig *pCfg) {
   if (cfgAddBool(pCfg, "queryUseNodeAllocator", tsQueryUseNodeAllocator, CFG_SCOPE_CLIENT) != 0) return -1;
   if (cfgAddBool(pCfg, "keepColumnName", tsKeepColumnName, CFG_SCOPE_CLIENT) != 0) return -1;
   if (cfgAddString(pCfg, "smlChildTableName", tsSmlChildTableName, CFG_SCOPE_CLIENT) != 0) return -1;
-  if (cfgAddString(pCfg, "smlAutoChildTableNameDelimiter", tsSmlAutoChildTableNameDelimiter, CFG_SCOPE_CLIENT) != 0) return -1;
+  if (cfgAddString(pCfg, "smlAutoChildTableNameDelimiter", tsSmlAutoChildTableNameDelimiter, CFG_SCOPE_CLIENT) != 0)
+    return -1;
   if (cfgAddString(pCfg, "smlTagName", tsSmlTagName, CFG_SCOPE_CLIENT) != 0) return -1;
   if (cfgAddString(pCfg, "smlTsDefaultName", tsSmlTsDefaultName, CFG_SCOPE_CLIENT) != 0) return -1;
   if (cfgAddBool(pCfg, "smlDot2Underline", tsSmlDot2Underline, CFG_SCOPE_CLIENT) != 0) return -1;
@@ -675,7 +678,8 @@ static int32_t taosAddServerCfg(SConfig *pCfg) {
   if (cfgAddInt64(pCfg, "checkpointInterval", tsStreamCheckpointInterval, 60, 1200, CFG_SCOPE_SERVER) != 0) return -1;
   if (cfgAddFloat(pCfg, "streamSinkDataRate", tsSinkDataRate, 0.1, 5, CFG_SCOPE_SERVER) != 0) return -1;
 
-  if (cfgAddInt32(pCfg, "cacheLazyLoadThreshold", tsCacheLazyLoadThreshold, 0, 100000, CFG_SCOPE_SERVER) != 0) return -1;
+  if (cfgAddInt32(pCfg, "cacheLazyLoadThreshold", tsCacheLazyLoadThreshold, 0, 100000, CFG_SCOPE_SERVER) != 0)
+    return -1;
 
   if (cfgAddString(pCfg, "lossyColumns", tsLossyColumns, CFG_SCOPE_SERVER) != 0) return -1;
   if (cfgAddFloat(pCfg, "fPrecision", tsFPrecision, 0.0f, 100000.0f, CFG_SCOPE_SERVER) != 0) return -1;
@@ -693,8 +697,11 @@ static int32_t taosAddServerCfg(SConfig *pCfg) {
   if (cfgAddString(pCfg, "s3Accesskey", tsS3AccessKey, CFG_SCOPE_SERVER) != 0) return -1;
   if (cfgAddString(pCfg, "s3Endpoint", tsS3Endpoint, CFG_SCOPE_SERVER) != 0) return -1;
   if (cfgAddString(pCfg, "s3BucketName", tsS3BucketName, CFG_SCOPE_SERVER) != 0) return -1;
-  if (cfgAddInt32(pCfg, "s3BlockSize", tsS3BlockSize, 2048, 1024 * 1024, CFG_SCOPE_SERVER) != 0) return -1;
+  if (cfgAddInt32(pCfg, "s3BlockSize", tsS3BlockSize, -100, 1024 * 1024, CFG_SCOPE_SERVER) != 0) return -1;
   if (cfgAddInt32(pCfg, "s3BlockCacheSize", tsS3BlockCacheSize, 4, 1024 * 1024, CFG_SCOPE_SERVER) != 0) return -1;
+  if (cfgAddInt32(pCfg, "s3PageCacheSize", tsS3PageCacheSize, 4, 1024 * 1024 * 1024, CFG_SCOPE_SERVER) != 0) return -1;
+  if (cfgAddInt32(pCfg, "s3UploadDelaySec", tsS3UploadDelaySec, 60 * 10, 60 * 60 * 24 * 30, CFG_SCOPE_SERVER) != 0)
+    return -1;
 
   // min free disk space used to check if the disk is full [50MB, 1GB]
   if (cfgAddInt64(pCfg, "minDiskFreeSize", tsMinDiskFreeSize, TFS_MIN_DISK_FREE_SIZE, 1024 * 1024 * 1024,
@@ -953,7 +960,8 @@ static int32_t taosSetClientCfg(SConfig *pCfg) {
     return -1;
   }
 
-  tstrncpy(tsSmlAutoChildTableNameDelimiter, cfgGetItem(pCfg, "smlAutoChildTableNameDelimiter")->str, TSDB_TABLE_NAME_LEN);
+  tstrncpy(tsSmlAutoChildTableNameDelimiter, cfgGetItem(pCfg, "smlAutoChildTableNameDelimiter")->str,
+           TSDB_TABLE_NAME_LEN);
   tstrncpy(tsSmlChildTableName, cfgGetItem(pCfg, "smlChildTableName")->str, TSDB_TABLE_NAME_LEN);
   tstrncpy(tsSmlTagName, cfgGetItem(pCfg, "smlTagName")->str, TSDB_COL_NAME_LEN);
   tstrncpy(tsSmlTsDefaultName, cfgGetItem(pCfg, "smlTsDefaultName")->str, TSDB_COL_NAME_LEN);
@@ -1125,6 +1133,8 @@ static int32_t taosSetServerCfg(SConfig *pCfg) {
 
   tsS3BlockSize = cfgGetItem(pCfg, "s3BlockSize")->i32;
   tsS3BlockCacheSize = cfgGetItem(pCfg, "s3BlockCacheSize")->i32;
+  tsS3PageCacheSize = cfgGetItem(pCfg, "s3PageCacheSize")->i32;
+  tsS3UploadDelaySec = cfgGetItem(pCfg, "s3UploadDelaySec")->i32;
 
   GRANT_CFG_GET;
   return 0;
@@ -1422,7 +1432,8 @@ int32_t taosApplyLocalCfg(SConfig *pCfg, char *name) {
       } else if (strcasecmp("smlChildTableName", name) == 0) {
         tstrncpy(tsSmlChildTableName, cfgGetItem(pCfg, "smlChildTableName")->str, TSDB_TABLE_NAME_LEN);
       } else if (strcasecmp("smlAutoChildTableNameDelimiter", name) == 0) {
-        tstrncpy(tsSmlAutoChildTableNameDelimiter, cfgGetItem(pCfg, "smlAutoChildTableNameDelimiter")->str, TSDB_TABLE_NAME_LEN);
+        tstrncpy(tsSmlAutoChildTableNameDelimiter, cfgGetItem(pCfg, "smlAutoChildTableNameDelimiter")->str,
+                 TSDB_TABLE_NAME_LEN);
       } else if (strcasecmp("smlTagName", name) == 0) {
         tstrncpy(tsSmlTagName, cfgGetItem(pCfg, "smlTagName")->str, TSDB_COL_NAME_LEN);
         //      } else if (strcasecmp("smlDataFormat", name) == 0) {
@@ -1717,6 +1728,20 @@ void taosCfgDynamicOptions(const char *option, const char *value) {
     return;
   }
 
+  if (strcasecmp(option, "s3PageCacheSize") == 0) {
+    int32_t newS3PageCacheSize = atoi(value);
+    uInfo("s3PageCacheSize set from %d to %d", tsS3PageCacheSize, newS3PageCacheSize);
+    tsS3PageCacheSize = newS3PageCacheSize;
+    return;
+  }
+
+  if (strcasecmp(option, "s3UploadDelaySec") == 0) {
+    int32_t newS3UploadDelaysec = atoi(value);
+    uInfo("s3UploadDelaySec set from %d to %d", tsS3UploadDelaySec, newS3UploadDelaysec);
+    tsS3UploadDelaySec = newS3UploadDelaysec;
+    return;
+  }
+
   if (strcasecmp(option, "ttlPushInterval") == 0) {
     int32_t newTtlPushInterval = atoi(value);
     uInfo("ttlPushInterval set from %d to %d", tsTtlPushIntervalSec, newTtlPushInterval);

source/dnode/mnode/impl/src/mndDnode.c

@@ -1228,6 +1228,70 @@ static int32_t mndProcessConfigDnodeReq(SRpcMsg *pReq) {
 
     strcpy(dcfgReq.config, "monitor");
     snprintf(dcfgReq.value, TSDB_DNODE_VALUE_LEN, "%d", flag);
+  } else if (strncasecmp(cfgReq.config, "s3blocksize", 11) == 0) {
+    int32_t optLen = strlen("s3blocksize");
+    int32_t flag = -1;
+    int32_t code = mndMCfgGetValInt32(&cfgReq, optLen, &flag);
+    if (code < 0) return code;
+
+    if (flag > 1024 * 1024) {
+      mError("dnode:%d, failed to config s3blocksize since value:%d. Valid range: [4, 1024 * 1024]", cfgReq.dnodeId,
+             flag);
+      terrno = TSDB_CODE_INVALID_CFG;
+      tFreeSMCfgDnodeReq(&cfgReq);
+      return -1;
+    }
+
+    strcpy(dcfgReq.config, "s3blocksize");
+    snprintf(dcfgReq.value, TSDB_DNODE_VALUE_LEN, "%d", flag);
+  } else if (strncasecmp(cfgReq.config, "s3blockcachesize", 16) == 0) {
+    int32_t optLen = strlen("s3blockcachesize");
+    int32_t flag = -1;
+    int32_t code = mndMCfgGetValInt32(&cfgReq, optLen, &flag);
+    if (code < 0) return code;
+
+    if (flag < 4 || flag > 1024 * 1024) {
+      mError("dnode:%d, failed to config s3BlockCacheSize since value:%d. Valid range: [4, 1024 * 1024]",
+             cfgReq.dnodeId, flag);
+      terrno = TSDB_CODE_INVALID_CFG;
+      tFreeSMCfgDnodeReq(&cfgReq);
+      return -1;
+    }
+
+    strcpy(dcfgReq.config, "s3blockcachesize");
+    snprintf(dcfgReq.value, TSDB_DNODE_VALUE_LEN, "%d", flag);
+  } else if (strncasecmp(cfgReq.config, "s3pagecachesize", 16) == 0) {
+    int32_t optLen = strlen("s3pagecachesize");
+    int32_t flag = -1;
+    int32_t code = mndMCfgGetValInt32(&cfgReq, optLen, &flag);
+    if (code < 0) return code;
+
+    if (flag < 4 || flag > 1024 * 1024 * 1024) {
+      mError("dnode:%d, failed to config s3PageCacheSize since value:%d. Valid range: [4, 1024 * 1024]", cfgReq.dnodeId,
+             flag);
+      terrno = TSDB_CODE_INVALID_CFG;
+      tFreeSMCfgDnodeReq(&cfgReq);
+      return -1;
+    }
+
+    strcpy(dcfgReq.config, "s3pagecachesize");
+    snprintf(dcfgReq.value, TSDB_DNODE_VALUE_LEN, "%d", flag);
+  } else if (strncasecmp(cfgReq.config, "s3uploaddelaysec", 16) == 0) {
+    int32_t optLen = strlen("s3uploaddelaysec");
+    int32_t flag = -1;
+    int32_t code = mndMCfgGetValInt32(&cfgReq, optLen, &flag);
+    if (code < 0) return code;
+
+    if (flag < 600 || flag > 60 * 60 * 24 * 30) {
+      mError("dnode:%d, failed to config s3UploadDelaySec since value:%d. Valid range: [600, 60 * 60 * 24 * 30]",
+             cfgReq.dnodeId, flag);
+      terrno = TSDB_CODE_INVALID_CFG;
+      tFreeSMCfgDnodeReq(&cfgReq);
+      return -1;
+    }
+
+    strcpy(dcfgReq.config, "s3uploaddelaysec");
+    snprintf(dcfgReq.value, TSDB_DNODE_VALUE_LEN, "%d", flag);
   } else if (strncasecmp(cfgReq.config, "ttlpushinterval", 14) == 0) {
     int32_t optLen = strlen("ttlpushinterval");
     int32_t flag = -1;

source/dnode/vnode/src/inc/tsdb.h

@@ -382,6 +382,8 @@ struct STsdb {
   TdThreadMutex        biMutex;
   SLRUCache           *bCache;
   TdThreadMutex        bMutex;
+  SLRUCache           *pgCache;
+  TdThreadMutex        pgMutex;
   struct STFileSystem *pFS;  // new
   SRocksCache          rCache;
 };
@@ -909,7 +911,9 @@ int32_t tsdbCacheGetBlockIdx(SLRUCache *pCache, SDataFReader *pFileReader, LRUHa
 int32_t tsdbBICacheRelease(SLRUCache *pCache, LRUHandle *h);
 
 int32_t tsdbCacheGetBlockS3(SLRUCache *pCache, STsdbFD *pFD, LRUHandle **handle);
-int32_t tsdbBCacheRelease(SLRUCache *pCache, LRUHandle *h);
+int32_t tsdbCacheGetPageS3(SLRUCache *pCache, STsdbFD *pFD, int64_t pgno, LRUHandle **handle);
+int32_t tsdbCacheSetPageS3(SLRUCache *pCache, STsdbFD *pFD, int64_t pgno, uint8_t *pPage);
+int32_t tsdbCacheRelease(SLRUCache *pCache, LRUHandle *h);
 
 int32_t tsdbCacheDeleteLastrow(SLRUCache *pCache, tb_uid_t uid, TSKEY eKey);
 int32_t tsdbCacheDeleteLast(SLRUCache *pCache, tb_uid_t uid, TSKEY eKey);

source/dnode/vnode/src/inc/vndCos.h

@@ -27,6 +27,8 @@ extern "C" {
 extern int8_t  tsS3Enabled;
 extern int32_t tsS3BlockSize;
 extern int32_t tsS3BlockCacheSize;
+extern int32_t tsS3PageCacheSize;
+extern int32_t tsS3UploadDelaySec;
 
 int32_t s3Init();
 void    s3CleanUp();

source/dnode/vnode/src/tsdb/tsdbCache.c

@@ -87,6 +87,41 @@ static void tsdbCloseBCache(STsdb *pTsdb) {
   }
 }
 
+static int32_t tsdbOpenPgCache(STsdb *pTsdb) {
+  int32_t code = 0;
+  // SLRUCache *pCache = taosLRUCacheInit(10 * 1024 * 1024, 0, .5);
+  int32_t szPage = pTsdb->pVnode->config.tsdbPageSize;
+
+  SLRUCache *pCache = taosLRUCacheInit((int64_t)tsS3PageCacheSize * szPage, 0, .5);
+  if (pCache == NULL) {
+    code = TSDB_CODE_OUT_OF_MEMORY;
+    goto _err;
+  }
+
+  taosLRUCacheSetStrictCapacity(pCache, false);
+
+  taosThreadMutexInit(&pTsdb->pgMutex, NULL);
+
+_err:
+  pTsdb->pgCache = pCache;
+  return code;
+}
+
+static void tsdbClosePgCache(STsdb *pTsdb) {
+  SLRUCache *pCache = pTsdb->pgCache;
+  if (pCache) {
+    int32_t elems = taosLRUCacheGetElems(pCache);
+    tsdbTrace("vgId:%d, elems: %d", TD_VID(pTsdb->pVnode), elems);
+    taosLRUCacheEraseUnrefEntries(pCache);
+    elems = taosLRUCacheGetElems(pCache);
+    tsdbTrace("vgId:%d, elems: %d", TD_VID(pTsdb->pVnode), elems);
+
+    taosLRUCacheCleanup(pCache);
+
+    taosThreadMutexDestroy(&pTsdb->bMutex);
+  }
+}
+
 #define ROCKS_KEY_LEN (sizeof(tb_uid_t) + sizeof(int16_t) + sizeof(int8_t))
 
 typedef struct {
@@ -1191,6 +1226,12 @@ int32_t tsdbOpenCache(STsdb *pTsdb) {
     goto _err;
   }
 
+  code = tsdbOpenPgCache(pTsdb);
+  if (code != TSDB_CODE_SUCCESS) {
+    code = TSDB_CODE_OUT_OF_MEMORY;
+    goto _err;
+  }
+
   code = tsdbOpenRocksCache(pTsdb);
   if (code != TSDB_CODE_SUCCESS) {
     code = TSDB_CODE_OUT_OF_MEMORY;
@@ -1221,6 +1262,7 @@ void tsdbCloseCache(STsdb *pTsdb) {
 
   tsdbCloseBICache(pTsdb);
   tsdbCloseBCache(pTsdb);
+  tsdbClosePgCache(pTsdb);
   tsdbCloseRocksCache(pTsdb);
 }
 
@@ -3057,7 +3099,6 @@ static int32_t tsdbCacheLoadBlockS3(STsdbFD *pFD, uint8_t **ppBlock) {
   }
   */
   int64_t block_offset = (pFD->blkno - 1) * tsS3BlockSize * pFD->szPage;
-  // int64_t size = 4096;
   code = s3GetObjectBlock(pFD->objName, block_offset, tsS3BlockSize * pFD->szPage, ppBlock);
   if (code != TSDB_CODE_SUCCESS) {
     // taosMemoryFree(pBlock);
@@ -3123,10 +3164,42 @@ int32_t tsdbCacheGetBlockS3(SLRUCache *pCache, STsdbFD *pFD, LRUHandle **handle)
   return code;
 }
 
-int32_t tsdbBCacheRelease(SLRUCache *pCache, LRUHandle *h) {
+int32_t tsdbCacheGetPageS3(SLRUCache *pCache, STsdbFD *pFD, int64_t pgno, LRUHandle **handle) {
   int32_t code = 0;
+  char    key[128] = {0};
+  int     keyLen = 0;
 
-  taosLRUCacheRelease(pCache, h, false);
+  getBCacheKey(pFD->fid, pFD->cid, pgno, key, &keyLen);
+  *handle = taosLRUCacheLookup(pCache, key, keyLen);
+
+  return code;
+}
+
+int32_t tsdbCacheSetPageS3(SLRUCache *pCache, STsdbFD *pFD, int64_t pgno, uint8_t *pPage) {
+  int32_t    code = 0;
+  char       key[128] = {0};
+  int        keyLen = 0;
+  LRUHandle *handle = NULL;
+
+  getBCacheKey(pFD->fid, pFD->cid, pgno, key, &keyLen);
+  taosThreadMutexLock(&pFD->pTsdb->pgMutex);
+  handle = taosLRUCacheLookup(pFD->pTsdb->pgCache, key, keyLen);
+  if (!handle) {
+    size_t              charge = pFD->szPage;
+    _taos_lru_deleter_t deleter = deleteBCache;
+    uint8_t            *pPg = taosMemoryMalloc(charge);
+    memcpy(pPg, pPage, charge);
+
+    LRUStatus status =
+        taosLRUCacheInsert(pCache, key, keyLen, pPg, charge, deleter, &handle, TAOS_LRU_PRIORITY_LOW, NULL);
+    if (status != TAOS_LRU_STATUS_OK) {
+      // ignore cache updating if not ok
+      // code = TSDB_CODE_OUT_OF_MEMORY;
+    }
+  }
+  taosThreadMutexUnlock(&pFD->pTsdb->pgMutex);
+
+  tsdbCacheRelease(pFD->pTsdb->pgCache, handle);
 
   return code;
 }

source/dnode/vnode/src/tsdb/tsdbCommit2.c

@@ -46,6 +46,7 @@ typedef struct {
     STFileSet *fset;
     TABLEID    tbid[1];
     bool       hasTSData;
+    bool       skipTsRow;
   } ctx[1];
 
   // reader
@@ -127,18 +128,18 @@ static int32_t tsdbCommitTSData(SCommitter2 *committer) {
         continue;
       }
     }
-
+    /*
     extern int8_t tsS3Enabled;
 
     int32_t nlevel = tfsGetLevel(committer->tsdb->pVnode->pTfs);
-    bool    skipRow = false;
+    committer->ctx->skipTsRow = false;
     if (tsS3Enabled && nlevel > 1 && committer->ctx->did.level == nlevel - 1) {
-      skipRow = true;
+      committer->ctx->skipTsRow = true;
     }
-
+    */
     int64_t ts = TSDBROW_TS(&row->row);
 
-    if (skipRow && ts <= committer->ctx->maxKey) {
+    if (committer->ctx->skipTsRow && ts <= committer->ctx->maxKey) {
       ts = committer->ctx->maxKey + 1;
     }
 
@@ -402,6 +403,32 @@ static int32_t tsdbCommitFileSetBegin(SCommitter2 *committer) {
   // reset nextKey
   committer->ctx->nextKey = TSKEY_MAX;
 
+  committer->ctx->skipTsRow = false;
+
+  extern int8_t  tsS3Enabled;
+  extern int32_t tsS3UploadDelaySec;
+  long           s3Size(const char *object_name);
+  int32_t        nlevel = tfsGetLevel(committer->tsdb->pVnode->pTfs);
+  committer->ctx->skipTsRow = false;
+  if (tsS3Enabled && nlevel > 1 && committer->ctx->fset) {
+    STFileObj *fobj = committer->ctx->fset->farr[TSDB_FTYPE_DATA];
+    if (fobj && fobj->f->did.level == nlevel - 1) {
+      // if exists on s3 or local mtime < committer->ctx->now - tsS3UploadDelay
+      const char *object_name = taosDirEntryBaseName((char *)fobj->fname);
+
+      if (taosCheckExistFile(fobj->fname)) {
+        int32_t mtime = 0;
+        taosStatFile(fobj->fname, NULL, &mtime, NULL);
+        if (mtime < committer->ctx->now - tsS3UploadDelaySec) {
+          committer->ctx->skipTsRow = true;
+        }
+      } else if (s3Size(object_name) > 0) {
+        committer->ctx->skipTsRow = true;
+      }
+    }
+    // new fset can be written with ts data
+  }
+
 _exit:
   if (code) {
     TSDB_ERROR_LOG(TD_VID(tsdb->pVnode), lino, code);

source/dnode/vnode/src/tsdb/tsdbFile2.c

@@ -303,6 +303,7 @@ bool tsdbIsSameTFile(const STFile *f1, const STFile *f2) {
   if (f1->did.id != f2->did.id) return false;
   if (f1->fid != f2->fid) return false;
   if (f1->cid != f2->cid) return false;
+  if (f1->s3flag != f2->s3flag) return false;
   return true;
 }
 

source/dnode/vnode/src/tsdb/tsdbFile2.h

@@ -58,6 +58,7 @@ int32_t tsdbTFileObjCmpr(const STFileObj **fobj1, const STFileObj **fobj2);
 struct STFile {
   tsdb_ftype_t type;
   SDiskID      did;  // disk id
+  int32_t      s3flag;
   int32_t      fid;  // file id
   int64_t      cid;  // commit id
   int64_t      size;

source/dnode/vnode/src/tsdb/tsdbReaderWriter.c

@@ -131,6 +131,7 @@ static int32_t tsdbWriteFilePage(STsdbFD *pFD) {
   }
 
   if (pFD->s3File) {
+    tsdbWarn("%s file: %s", __func__, pFD->path);
     return code;
   }
   if (pFD->pgno > 0) {
@@ -177,7 +178,7 @@ static int32_t tsdbReadFilePage(STsdbFD *pFD, int64_t pgno) {
     pFD->blkno = (pgno + tsS3BlockSize - 1) / tsS3BlockSize;
     code = tsdbCacheGetBlockS3(pFD->pTsdb->bCache, pFD, &handle);
     if (code != TSDB_CODE_SUCCESS || handle == NULL) {
-      tsdbBCacheRelease(pFD->pTsdb->bCache, handle);
+      tsdbCacheRelease(pFD->pTsdb->bCache, handle);
       if (code == TSDB_CODE_SUCCESS && !handle) {
         code = TSDB_CODE_OUT_OF_MEMORY;
       }
@@ -189,7 +190,7 @@ static int32_t tsdbReadFilePage(STsdbFD *pFD, int64_t pgno) {
     int64_t blk_offset = (pFD->blkno - 1) * tsS3BlockSize * pFD->szPage;
     memcpy(pFD->pBuf, pBlock + (offset - blk_offset), pFD->szPage);
 
-    tsdbBCacheRelease(pFD->pTsdb->bCache, handle);
+    tsdbCacheRelease(pFD->pTsdb->bCache, handle);
   } else {
     // seek
     int64_t n = taosLSeekFile(pFD->pFD, offset, SEEK_SET);
@@ -253,7 +254,7 @@ _exit:
   return code;
 }
 
-int32_t tsdbReadFile(STsdbFD *pFD, int64_t offset, uint8_t *pBuf, int64_t size) {
+static int32_t tsdbReadFileImp(STsdbFD *pFD, int64_t offset, uint8_t *pBuf, int64_t size) {
   int32_t code = 0;
   int64_t n = 0;
   int64_t fOffset = LOGIC_TO_FILE_OFFSET(offset, pFD->szPage);
@@ -282,10 +283,122 @@ _exit:
   return code;
 }
 
+static int32_t tsdbReadFileS3(STsdbFD *pFD, int64_t offset, uint8_t *pBuf, int64_t size) {
+  int32_t code = 0;
+  int64_t n = 0;
+  int32_t szPgCont = PAGE_CONTENT_SIZE(pFD->szPage);
+  int64_t fOffset = LOGIC_TO_FILE_OFFSET(offset, pFD->szPage);
+  int64_t pgno = OFFSET_PGNO(fOffset, pFD->szPage);
+  int64_t bOffset = fOffset % pFD->szPage;
+
+  ASSERT(bOffset < szPgCont);
+
+  // 1, find pgnoStart & pgnoEnd to fetch from s3, if all pgs are local, no need to fetch
+  // 2, fetch pgnoStart ~ pgnoEnd from s3
+  // 3, store pgs to pcache & last pg to pFD->pBuf
+  // 4, deliver pgs to [pBuf, pBuf + size)
+
+  while (n < size) {
+    if (pFD->pgno != pgno) {
+      LRUHandle *handle = NULL;
+      code = tsdbCacheGetPageS3(pFD->pTsdb->pgCache, pFD, pgno, &handle);
+      if (code != TSDB_CODE_SUCCESS) {
+        if (handle) {
+          tsdbCacheRelease(pFD->pTsdb->pgCache, handle);
+        }
+        goto _exit;
+      }
+
+      if (!handle) {
+        break;
+      }
+
+      uint8_t *pPage = (uint8_t *)taosLRUCacheValue(pFD->pTsdb->pgCache, handle);
+      memcpy(pFD->pBuf, pPage, pFD->szPage);
+      tsdbCacheRelease(pFD->pTsdb->pgCache, handle);
+
+      // check
+      if (pgno > 1 && !taosCheckChecksumWhole(pFD->pBuf, pFD->szPage)) {
+        code = TSDB_CODE_FILE_CORRUPTED;
+        goto _exit;
+      }
+
+      pFD->pgno = pgno;
+    }
+
+    int64_t nRead = TMIN(szPgCont - bOffset, size - n);
+    memcpy(pBuf + n, pFD->pBuf + bOffset, nRead);
+
+    n += nRead;
+    pgno++;
+    bOffset = 0;
+  }
+
+  if (n < size) {
+    // 2, retrieve pgs from s3
+    uint8_t *pBlock = NULL;
+    int64_t  retrieve_offset = PAGE_OFFSET(pgno, pFD->szPage);
+    int64_t  pgnoEnd = pgno - 1 + (size - n + szPgCont - 1) / szPgCont;
+    int64_t  retrieve_size = (pgnoEnd - pgno + 1) * pFD->szPage;
+    code = s3GetObjectBlock(pFD->objName, retrieve_offset, retrieve_size, &pBlock);
+    if (code != TSDB_CODE_SUCCESS) {
+      goto _exit;
+    }
+
+    // 3, Store Pages in Cache
+    int nPage = pgnoEnd - pgno + 1;
+    for (int i = 0; i < nPage; ++i) {
+      tsdbCacheSetPageS3(pFD->pTsdb->pgCache, pFD, pgno, pBlock + i * pFD->szPage);
+
+      memcpy(pFD->pBuf, pBlock + i * pFD->szPage, pFD->szPage);
+
+      // check
+      if (pgno > 1 && !taosCheckChecksumWhole(pFD->pBuf, pFD->szPage)) {
+        code = TSDB_CODE_FILE_CORRUPTED;
+        goto _exit;
+      }
+
+      pFD->pgno = pgno;
+
+      int64_t nRead = TMIN(szPgCont - bOffset, size - n);
+      memcpy(pBuf + n, pFD->pBuf + bOffset, nRead);
+
+      n += nRead;
+      pgno++;
+      bOffset = 0;
+    }
+
+    taosMemoryFree(pBlock);
+  }
+
+_exit:
+  return code;
+}
+
+int32_t tsdbReadFile(STsdbFD *pFD, int64_t offset, uint8_t *pBuf, int64_t size) {
+  int32_t code = 0;
+  if (!pFD->pFD) {
+    code = tsdbOpenFileImpl(pFD);
+    if (code) {
+      goto _exit;
+    }
+  }
+
+  if (pFD->s3File && tsS3BlockSize < 0) {
+    return tsdbReadFileS3(pFD, offset, pBuf, size);
+  } else {
+    return tsdbReadFileImp(pFD, offset, pBuf, size);
+  }
+
+_exit:
+  return code;
+}
+
 int32_t tsdbFsyncFile(STsdbFD *pFD) {
   int32_t code = 0;
 
   if (pFD->s3File) {
+    tsdbWarn("%s file: %s", __func__, pFD->path);
     return code;
   }
   code = tsdbWriteFilePage(pFD);

source/dnode/vnode/src/tsdb/tsdbRetention.c

@@ -206,6 +206,8 @@ static int32_t tsdbMigrateDataFileS3(SRTNer *rtner, const STFileObj *fobj, const
           },
   };
 
+  op.nf.s3flag = true;
+
   code = TARRAY2_APPEND(rtner->fopArr, op);
   TSDB_CHECK_CODE(code, lino, _exit);
 
@@ -322,27 +324,40 @@ static int32_t tsdbDoRetentionOnFileSet(SRTNer *rtner, STFileSet *fset) {
     for (int32_t ftype = 0; ftype < TSDB_FTYPE_MAX && (fobj = fset->farr[ftype], 1); ++ftype) {
       if (fobj == NULL) continue;
 
-      if (fobj->f->did.level == did.level) continue;
-
       int32_t nlevel = tfsGetLevel(rtner->tsdb->pVnode->pTfs);
+
+      if (fobj->f->did.level == did.level) {
+        if (tsS3Enabled && nlevel > 1 && TSDB_FTYPE_DATA == ftype && did.level == nlevel - 1 &&
+            taosCheckExistFile(fobj->fname)) {
+          int32_t mtime = 0;
+          taosStatFile(fobj->fname, NULL, &mtime, NULL);
+          if (mtime < rtner->now - tsS3UploadDelaySec) {
+            code = tsdbMigrateDataFileS3(rtner, fobj, &did);
+            TSDB_CHECK_CODE(code, lino, _exit);
+          }
+        }
+
+        continue;
+      }
+      /*
     if (tsS3Enabled && nlevel > 1 && TSDB_FTYPE_DATA == ftype && did.level == nlevel - 1) {
       code = tsdbMigrateDataFileS3(rtner, fobj, &did);
       TSDB_CHECK_CODE(code, lino, _exit);
     } else {
+
       if (tsS3Enabled) {
         int64_t fsize = 0;
         if (taosStatFile(fobj->fname, &fsize, NULL, NULL) < 0) {
           code = TAOS_SYSTEM_ERROR(terrno);
-            tsdbError("vgId:%d %s failed since file:%s stat failed, reason:%s", TD_VID(rtner->tsdb->pVnode), __func__,
+          tsdbError("vgId:%d %s failed since file:%s stat failed, reason:%s", TD_VID(rtner->tsdb->pVnode),
-                      fobj->fname, tstrerror(code));
+      __func__, fobj->fname, tstrerror(code)); TSDB_CHECK_CODE(code, lino, _exit);
-            TSDB_CHECK_CODE(code, lino, _exit);
         }
         s3EvictCache(fobj->fname, fsize * 2);
       }
-
+      */
       code = tsdbDoMigrateFileObj(rtner, fobj, &did);
       TSDB_CHECK_CODE(code, lino, _exit);
-      }
+      //}
     }
 
     // stt

source/dnode/vnode/src/tsdb/tsdbWrite.c

@@ -80,8 +80,8 @@ int tsdbScanAndConvertSubmitMsg(STsdb *pTsdb, SSubmitReq2 *pMsg) {
   TSKEY         minKey = now - tsTickPerMin[pCfg->precision] * pCfg->keep2;
   TSKEY         maxKey = tsMaxKeyByPrecision[pCfg->precision];
   int32_t       size = taosArrayGetSize(pMsg->aSubmitTbData);
+  /*
   int32_t nlevel = tfsGetLevel(pTsdb->pVnode->pTfs);
-
   if (nlevel > 1 && tsS3Enabled) {
     if (nlevel == 3) {
       minKey = now - tsTickPerMin[pCfg->precision] * pCfg->keep1;
@@ -89,7 +89,7 @@ int tsdbScanAndConvertSubmitMsg(STsdb *pTsdb, SSubmitReq2 *pMsg) {
       minKey = now - tsTickPerMin[pCfg->precision] * pCfg->keep0;
     }
   }
-
+  */
   for (int32_t i = 0; i < size; ++i) {
     SSubmitTbData *pData = TARRAY_GET_ELEM(pMsg->aSubmitTbData, i);
     if (pData->flags & SUBMIT_REQ_COLUMN_DATA_FORMAT) {

source/dnode/vnode/src/vnd/vnodeSvr.c

@@ -13,14 +13,14 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
+#include "audit.h"
 #include "tencode.h"
 #include "tmsg.h"
+#include "tstrbuild.h"
 #include "vnd.h"
 #include "vndCos.h"
 #include "vnode.h"
 #include "vnodeInt.h"
-#include "audit.h"
-#include "tstrbuild.h"
 
 static int32_t vnodeProcessCreateStbReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp);
 static int32_t vnodeProcessAlterStbReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp);
@@ -263,8 +263,9 @@ static int32_t vnodePreProcessSubmitTbData(SVnode *pVnode, SDecoder *pCoder, int
     now *= 1000000;
   }
 
-  int32_t nlevel = tfsGetLevel(pVnode->pTfs);
   int32_t keep = pVnode->config.tsdbCfg.keep2;
+  /*
+  int32_t nlevel = tfsGetLevel(pVnode->pTfs);
   if (nlevel > 1 && tsS3Enabled) {
     if (nlevel == 3) {
       keep = pVnode->config.tsdbCfg.keep1;
@@ -272,6 +273,7 @@ static int32_t vnodePreProcessSubmitTbData(SVnode *pVnode, SDecoder *pCoder, int
       keep = pVnode->config.tsdbCfg.keep0;
     }
   }
+  */
 
   TSKEY minKey = now - tsTickPerMin[pVnode->config.tsdbCfg.precision] * keep;
   TSKEY maxKey = tsMaxKeyByPrecision[pVnode->config.tsdbCfg.precision];
@@ -1518,7 +1520,6 @@ static int32_t vnodeProcessSubmitReq(SVnode *pVnode, int64_t ver, void *pReq, in
       int32_t nRow = TARRAY_SIZE(pSubmitTbData->aRowP);
       SRow  **aRow = (SRow **)TARRAY_DATA(pSubmitTbData->aRowP);
       for (int32_t iRow = 0; iRow < nRow; ++iRow) {
-
         if (aRow[iRow]->ts < minKey || aRow[iRow]->ts > maxKey || (iRow > 0 && aRow[iRow]->ts <= aRow[iRow - 1]->ts)) {
           code = TSDB_CODE_INVALID_MSG;
           vError("vgId:%d %s failed since %s, version:%" PRId64, TD_VID(pVnode), __func__, tstrerror(code), ver);