]>
Commit | Line | Data |
---|---|---|
7c673cae | 1 | // Copyright (c) 2011-present, Facebook, Inc. All rights reserved. |
11fdf7f2 TL |
2 | // This source code is licensed under both the GPLv2 (found in the |
3 | // COPYING file in the root directory) and Apache 2.0 License | |
4 | // (found in the LICENSE.Apache file in the root directory). | |
7c673cae FG |
5 | |
6 | #pragma once | |
7 | #ifndef ROCKSDB_LITE | |
8 | ||
9 | #include <map> | |
10 | #include <string> | |
11 | #include <vector> | |
12 | ||
13 | #include "rocksdb/db.h" | |
14 | ||
15 | namespace rocksdb { | |
16 | ||
17 | // Date tiered database is a wrapper of DB that implements | |
18 | // a simplified DateTieredCompactionStrategy by using multiple column famillies | |
19 | // as time windows. | |
20 | // | |
21 | // DateTieredDB provides an interface similar to DB, but it assumes that user | |
22 | // provides keys with last 8 bytes encoded as timestamp in seconds. DateTieredDB | |
23 | // is assigned with a TTL to declare when data should be deleted. | |
24 | // | |
25 | // DateTieredDB hides column families layer from standard RocksDB instance. It | |
26 | // uses multiple column families to manage time series data, each containing a | |
27 | // specific range of time. Column families are named by its maximum possible | |
28 | // timestamp. A column family is created automatically when data newer than | |
29 | // latest timestamp of all existing column families. The time range of a column | |
30 | // family is configurable by `column_family_interval`. By doing this, we | |
31 | // guarantee that compaction will only happen in a column family. | |
32 | // | |
33 | // DateTieredDB is assigned with a TTL. When all data in a column family are | |
34 | // expired (CF_Timestamp <= CUR_Timestamp - TTL), we directly drop the whole | |
35 | // column family. | |
36 | // | |
37 | // TODO(jhli): This is only a simplified version of DTCS. In a complete DTCS, | |
38 | // time windows can be merged over time, so that older time windows will have | |
39 | // larger time range. Also, compaction are executed only for adjacent SST files | |
40 | // to guarantee there is no time overlap between SST files. | |
41 | ||
42 | class DateTieredDB { | |
43 | public: | |
44 | // Open a DateTieredDB whose name is `dbname`. | |
45 | // Similar to DB::Open(), created database object is stored in dbptr. | |
46 | // | |
47 | // Two parameters can be configured: `ttl` to specify the length of time that | |
48 | // keys should exist in the database, and `column_family_interval` to specify | |
49 | // the time range of a column family interval. | |
50 | // | |
51 | // Open a read only database if read only is set as true. | |
52 | // TODO(jhli): Should use an option object that includes ttl and | |
53 | // column_family_interval. | |
54 | static Status Open(const Options& options, const std::string& dbname, | |
55 | DateTieredDB** dbptr, int64_t ttl, | |
56 | int64_t column_family_interval, bool read_only = false); | |
57 | ||
58 | explicit DateTieredDB() {} | |
59 | ||
60 | virtual ~DateTieredDB() {} | |
61 | ||
62 | // Wrapper for Put method. Similar to DB::Put(), but column family to be | |
63 | // inserted is decided by the timestamp in keys, i.e. the last 8 bytes of user | |
64 | // key. If key is already obsolete, it will not be inserted. | |
65 | // | |
66 | // When client put a key value pair in DateTieredDB, it assumes last 8 bytes | |
67 | // of keys are encoded as timestamp. Timestamp is a 64-bit signed integer | |
68 | // encoded as the number of seconds since 1970-01-01 00:00:00 (UTC) (Same as | |
69 | // Env::GetCurrentTime()). Timestamp should be encoded in big endian. | |
70 | virtual Status Put(const WriteOptions& options, const Slice& key, | |
71 | const Slice& val) = 0; | |
72 | ||
73 | // Wrapper for Get method. Similar to DB::Get() but column family is decided | |
74 | // by timestamp in keys. If key is already obsolete, it will not be found. | |
75 | virtual Status Get(const ReadOptions& options, const Slice& key, | |
76 | std::string* value) = 0; | |
77 | ||
78 | // Wrapper for Delete method. Similar to DB::Delete() but column family is | |
79 | // decided by timestamp in keys. If key is already obsolete, return NotFound | |
80 | // status. | |
81 | virtual Status Delete(const WriteOptions& options, const Slice& key) = 0; | |
82 | ||
83 | // Wrapper for KeyMayExist method. Similar to DB::KeyMayExist() but column | |
84 | // family is decided by timestamp in keys. Return false when key is already | |
85 | // obsolete. | |
86 | virtual bool KeyMayExist(const ReadOptions& options, const Slice& key, | |
87 | std::string* value, bool* value_found = nullptr) = 0; | |
88 | ||
89 | // Wrapper for Merge method. Similar to DB::Merge() but column family is | |
90 | // decided by timestamp in keys. | |
91 | virtual Status Merge(const WriteOptions& options, const Slice& key, | |
92 | const Slice& value) = 0; | |
93 | ||
94 | // Create an iterator that hides low level details. This iterator internally | |
95 | // merge results from all active time series column families. Note that | |
96 | // column families are not deleted until all data are obsolete, so this | |
97 | // iterator can possibly access obsolete key value pairs. | |
98 | virtual Iterator* NewIterator(const ReadOptions& opts) = 0; | |
99 | ||
100 | // Explicitly drop column families in which all keys are obsolete. This | |
101 | // process is also inplicitly done in Put() operation. | |
102 | virtual Status DropObsoleteColumnFamilies() = 0; | |
103 | ||
104 | static const uint64_t kTSLength = sizeof(int64_t); // size of timestamp | |
105 | }; | |
106 | ||
107 | } // namespace rocksdb | |
108 | #endif // ROCKSDB_LITE |