Permalink
Apr 23, 2015
Apr 23, 2015
Nov 1, 2017
Apr 23, 2015
Sep 30, 2016
Sep 30, 2016
Dec 28, 2016
Sep 30, 2016
Sep 30, 2016
Sep 30, 2016
Sep 30, 2016
Sep 30, 2016
Apr 23, 2015
Feb 13, 2017
Apr 23, 2015
Feb 13, 2017
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Sep 30, 2016
Apr 23, 2015
Oct 2, 2016
Apr 23, 2015
Oct 3, 2016
Apr 23, 2015
Oct 1, 2016
Apr 23, 2015
May 22, 2015
Apr 23, 2015
May 22, 2015
Apr 23, 2015
Jun 2, 2015
Oct 13, 2016
May 29, 2015
May 29, 2015
Apr 23, 2015
Apr 23, 2015
Sep 30, 2016
May 25, 2015
Sep 30, 2016
Apr 23, 2015
May 29, 2015
Apr 23, 2015
Sep 30, 2016
Apr 23, 2015
May 22, 2015
Apr 23, 2015
May 22, 2015
Apr 23, 2015
May 22, 2015
May 22, 2015
Apr 23, 2015
Sep 30, 2016
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
May 22, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
May 30, 2015
May 21, 2015
Apr 23, 2015
May 25, 2015
Apr 23, 2015
May 29, 2015
Jun 2, 2015
Apr 23, 2015
May 29, 2015
Apr 23, 2015
Sep 30, 2016
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Oct 13, 2016
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Jun 2, 2015
Apr 23, 2015
Apr 23, 2015
Jun 2, 2015
Jun 2, 2015
Jun 2, 2015
May 29, 2015
Jun 2, 2015
Apr 23, 2015
May 28, 2015
Apr 23, 2015
Jun 2, 2015
Apr 23, 2015
Sep 30, 2016
Apr 23, 2015
Sep 30, 2016
Apr 23, 2015
Apr 23, 2015
Sep 30, 2016
Apr 23, 2015
Sep 30, 2016
Apr 23, 2015
Sep 30, 2016
Apr 23, 2015
Apr 23, 2015
May 30, 2015
May 30, 2015
Apr 23, 2015
Oct 3, 2016
Apr 23, 2015
Oct 3, 2016
Apr 23, 2015
Oct 3, 2016
Apr 23, 2015
Oct 3, 2016
Apr 23, 2015
Oct 3, 2016
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Jun 17, 2015
Jun 17, 2015
Jun 17, 2015
Jun 17, 2015
Jun 17, 2015
Jun 17, 2015
Jun 17, 2015
Oct 1, 2016
Jan 26, 2017
Jun 17, 2015
Oct 1, 2016
Oct 1, 2016
Oct 2, 2016
Aug 29, 2015
Aug 29, 2015
Jul 10, 2016
Aug 29, 2015
Aug 29, 2015
Aug 29, 2015
Jun 17, 2015
Jun 17, 2015
Jun 17, 2015
Jun 17, 2015
Jun 17, 2015
Apr 23, 2015
Apr 23, 2015
Oct 1, 2016
Aug 28, 2017
Apr 23, 2015
Oct 3, 2017
Apr 23, 2015
Oct 24, 2017
Oct 3, 2017
Apr 23, 2015
Apr 23, 2015
Sep 30, 2016
Sep 30, 2016
Sep 30, 2016
Sep 30, 2016
Sep 30, 2016
Sep 30, 2016
Aug 28, 2017
Sep 30, 2016
Newer
100644
1475 lines (1213 sloc)
71.3 KB
1
# About
3
This document is an updated version of the original design documents
4
by Spencer Kimball from early 2014. It may not always be completely up to date.
5
For a more approachable explanation of how CockroachDB works, consider reading
6
the [Architecture docs](https://www.cockroachlabs.com/docs/stable/architecture/overview.html).
7
8
# Overview
9
10
CockroachDB is a distributed SQL database. The primary design goals
11
are **scalability**, **strong consistency** and **survivability**
15
**homogeneous deployment** (one binary) with minimal configuration and
16
no required external dependencies.
17
18
The entry point for database clients is the SQL interface. Every node
19
in a CockroachDB cluster can act as a client SQL gateway. A SQL
20
gateway transforms and executes client SQL statements to key-value
21
(KV) operations, which the gateway distributes across the cluster as
22
necessary and returns results to the client. CockroachDB implements a
23
**single, monolithic sorted map** from key to value where both keys
24
and values are byte strings.
25
26
The KV map is logically composed of smaller segments of the keyspace called
27
ranges. Each range is backed by data stored in a local KV storage engine (we
28
use [RocksDB](http://rocksdb.org/), a variant of
29
[LevelDB](https://github.com/google/leveldb)). Range data is replicated to a
30
configurable number of additional CockroachDB nodes. Ranges are merged and
31
split to maintain a target size, by default `64M`. The relatively small size
32
facilitates quick repair and rebalancing to address node failures, new capacity
33
and even read/write load. However, the size must be balanced against the
34
pressure on the system from having more ranges to manage.
35
36
CockroachDB achieves horizontally scalability:
37
- adding more nodes increases the capacity of the cluster by the
38
amount of storage on each node (divided by a configurable
39
replication factor), theoretically up to 4 exabytes (4E) of logical
40
data;
41
- client queries can be sent to any node in the cluster, and queries
42
can operate independently (w/o conflicts), meaning that overall
43
throughput is a linear factor of the number of nodes in the cluster.
48
- uses a distributed consensus protocol for synchronous replication of
49
data in each key value range. We’ve chosen to use the [Raft
50
consensus algorithm](https://raftconsensus.github.io); all consensus
51
state is stored in RocksDB.
52
- single or batched mutations to a single range are mediated via the
53
range's Raft instance. Raft guarantees ACID semantics.
54
- logical mutations which affect multiple ranges employ distributed
55
transactions for ACID semantics. CockroachDB uses an efficient
56
**non-locking distributed commit** protocol.
57
58
CockroachDB achieves survivability:
59
- range replicas can be co-located within a single datacenter for low
60
latency replication and survive disk or machine failures. They can
62
- range replicas can be located in datacenters spanning increasingly
63
disparate geographies to survive ever-greater failure scenarios from
64
datacenter power or networking loss to regional power failures
66
Japan }`, `{ Ireland, US-East, US-West}`, `{ Ireland, US-East,
67
US-West, Japan, Australia }`).
68
69
CockroachDB provides [snapshot
70
isolation](http://en.wikipedia.org/wiki/Snapshot_isolation) (SI) and
71
serializable snapshot isolation (SSI) semantics, allowing **externally
72
consistent, lock-free reads and writes**--both from a historical snapshot
73
timestamp and from the current wall clock time. SI provides lock-free reads
74
and writes but still allows write skew. SSI eliminates write skew, but
75
introduces a performance hit in the case of a contentious system. SSI is the
76
default isolation; clients must consciously decide to trade correctness for
77
performance. CockroachDB implements [a limited form of linearizability
78
](#strict-serializability-linearizability), providing ordering for any
79
observer or chain of observers.
80
81
Similar to
82
[Spanner](http://static.googleusercontent.com/media/research.google.com/en/us/archive/spanner-osdi2012.pdf)
84
This allows replication factor, storage device type, and/or datacenter
85
location to be chosen to optimize performance and/or availability.
86
Unlike Spanner, zones are monolithic and don’t allow movement of fine
87
grained data on the level of entity groups.
88
89
# Architecture
90
92
abstraction is the SQL layer (currently unspecified in this document).
93
It depends directly on the [*SQL layer*](#sql),
94
which provides familiar relational concepts
95
such as schemas, tables, columns, and indexes. The SQL layer
96
in turn depends on the [distributed key value store](#key-value-api),
97
which handles the details of range addressing to provide the abstraction
98
of a single, monolithic key value store. The distributed KV store
99
communicates with any number of physical cockroach nodes. Each node
100
contains one or more stores, one per physical device.
101
102
103
104
Each store contains potentially many ranges, the lowest-level unit of
105
key-value data. Ranges are replicated using the Raft consensus protocol.
106
The diagram below is a blown up version of stores from four of the five
107
nodes in the previous diagram. Each range is replicated three ways using
108
raft. The color coding shows associated range replicas.
109
110
111
112
Each physical node exports two RPC-based key value APIs: one for
113
external clients and one for internal clients (exposing sensitive
114
operational features). Both services accept batches of requests and
115
return batches of responses. Nodes are symmetric in capabilities and
116
exported interfaces; each has the same binary and may assume any
117
role.
118
119
Nodes and the ranges they provide access to can be arranged with various
120
physical network topologies to make trade offs between reliability and
121
performance. For example, a triplicated (3-way replica) range could have
122
each replica located on different:
123
124
- disks within a server to tolerate disk failures.
125
- servers within a rack to tolerate server failures.
126
- servers on different racks within a datacenter to tolerate rack power/network failures.
127
- servers in different datacenters to tolerate large scale network or power outages.
128
129
Up to `F` failures can be tolerated, where the total number of replicas `N = 2F + 1` (e.g. with 3x replication, one failure can be tolerated; with 5x replication, two failures, and so on).
130
131
# Keys
132
133
Cockroach keys are arbitrary byte arrays. Keys come in two flavors:
134
system keys and table data keys. System keys are used by Cockroach for
135
internal data structures and metadata. Table data keys contain SQL
136
table data (as well as index data). System and table data keys are
137
prefixed in such a way that all system keys sort before any table data
138
keys.
139
140
System keys come in several subtypes:
141
142
- **Global** keys store cluster-wide data such as the "meta1" and
143
"meta2" keys as well as various other system-wide keys such as the
144
node and store ID allocators.
145
- **Store local** keys are used for unreplicated store metadata
146
(e.g. the `StoreIdent` structure). "Unreplicated" indicates that
147
these values are not replicated across multiple stores because the
148
data they hold is tied to the lifetime of the store they are
149
present on.
150
- **Range local** keys store range metadata that is associated with a
151
global key. Range local keys have a special prefix followed by a
152
global key and a special suffix. For example, transaction records
153
are range local keys which look like:
154
`\x01k<global-key>txn-<txnID>`.
155
- **Replicated Range ID local** keys store range metadata that is
156
present on all of the replicas for a range. These keys are updated
157
via Raft operations. Examples include the range lease state and
158
abort cache entries.
159
- **Unreplicated Range ID local** keys store range metadata that is
160
local to a replica. The primary examples of such keys are the Raft
161
state and Raft log.
162
163
Table data keys are used to store all SQL data. Table data keys
164
contain internal structure as described in the section on [mapping
165
data between the SQL model and
166
KV](#data-mapping-between-the-sql-model-and-kv).
167
168
# Versioned Values
169
170
Cockroach maintains historical versions of values by storing them with
171
associated commit timestamps. Reads and scans can specify a snapshot
172
time to return the most recent writes prior to the snapshot timestamp.
173
Older versions of values are garbage collected by the system during
174
compaction according to a user-specified expiration interval. In order
175
to support long-running scans (e.g. for MapReduce), all versions have a
176
minimum expiration.
177
178
Versioned values are supported via modifications to RocksDB to record
179
commit timestamps and GC expirations per key.
180
181
# Lock-Free Distributed Transactions
182
183
Cockroach provides distributed transactions without locks. Cockroach
184
transactions support two isolation levels:
185
186
- snapshot isolation (SI) and
187
- *serializable* snapshot isolation (SSI).
188
189
*SI* is simple to implement, highly performant, and correct for all but a
190
handful of anomalous conditions (e.g. write skew). *SSI* requires just a touch
191
more complexity, is still highly performant (less so with contention), and has
192
no anomalous conditions. Cockroach’s SSI implementation is based on ideas from
193
the literature and some possibly novel insights.
194
195
SSI is the default level, with SI provided for application developers
196
who are certain enough of their need for performance and the absence of
197
write skew conditions to consciously elect to use it. In a lightly
198
contended system, our implementation of SSI is just as performant as SI,
199
requiring no locking or additional writes. With contention, our
200
implementation of SSI still requires no locking, but will end up
201
aborting more transactions. Cockroach’s SI and SSI implementations
202
prevent starvation scenarios even for arbitrarily long transactions.
203
204
See the [Cahill paper](https://drive.google.com/file/d/0B9GCVTp_FHJIcEVyZVdDWEpYYXVVbFVDWElrYUV0NHFhU2Fv/edit?usp=sharing)
205
for one possible implementation of SSI. This is another [great paper](http://cs.yale.edu/homes/thomson/publications/calvin-sigmod12.pdf).
206
For a discussion of SSI implemented by preventing read-write conflicts
207
(in contrast to detecting them, called write-snapshot isolation), see
208
the [Yabandeh paper](https://drive.google.com/file/d/0B9GCVTp_FHJIMjJ2U2t6aGpHLTFUVHFnMTRUbnBwc2pLa1RN/edit?usp=sharing),
209
which is the source of much inspiration for Cockroach’s SSI.
210
211
Both SI and SSI require that the outcome of reads must be preserved, i.e.
212
a write of a key at a lower timestamp than a previous read must not succeed. To
213
this end, each range maintains a bounded *in-memory* cache from key range to
214
the latest timestamp at which it was read.
215
216
Most updates to this *timestamp cache* correspond to keys being read, though
217
the timestamp cache also protects the outcome of some writes (notably range
218
deletions) which consequently must also populate the cache. The cache’s entries
219
are evicted oldest timestamp first, updating the low water mark of the cache
220
appropriately.
221
222
Each Cockroach transaction is assigned a random priority and a
223
"candidate timestamp" at start. The candidate timestamp is the
224
provisional timestamp at which the transaction will commit, and is
225
chosen as the current clock time of the node coordinating the
226
transaction. This means that a transaction without conflicts will
227
usually commit with a timestamp that, in absolute time, precedes the
228
actual work done by that transaction.
229
230
In the course of coordinating a transaction between one or more
231
distributed nodes, the candidate timestamp may be increased, but will
232
never be decreased. The core difference between the two isolation levels
233
SI and SSI is that the former allows the transaction's candidate
234
timestamp to increase and the latter does not.
235
239
in the [Hybrid Logical Clock paper](http://www.cse.buffalo.edu/tech-reports/2014-04.pdf).
240
HLC time uses timestamps which are composed of a physical component (thought of
241
as and always close to local wall time) and a logical component (used to
242
distinguish between events with the same physical component). It allows us to
243
track causality for related events similar to vector clocks, but with less
244
overhead. In practice, it works much like other logical clocks: When events
245
are received by a node, it informs the local HLC about the timestamp supplied
246
with the event by the sender, and when events are sent a timestamp generated by
247
the local HLC is attached.
248
249
For a more in depth description of HLC please read the paper. Our
250
implementation is [here](https://github.com/cockroachdb/cockroach/blob/master/pkg/util/hlc/hlc.go).
251
252
Cockroach picks a Timestamp for a transaction using HLC time. Throughout this
253
document, *timestamp* always refers to the HLC time which is a singleton
254
on each node. The HLC is updated by every read/write event on the node, and
256
from another node is not only used to version the operation, but also updates
257
the HLC on the node. This is useful in guaranteeing that all data read/written
258
on a node is at a timestamp < next HLC time.
259
263
264
1. Start the transaction by selecting a range which is likely to be
265
heavily involved in the transaction and writing a new transaction
266
record to a reserved area of that range with state "PENDING". In
267
parallel write an "intent" value for each datum being written as part
268
of the transaction. These are normal MVCC values, with the addition of
269
a special flag (i.e. “intent”) indicating that the value may be
272
is stored with intent values. The txn id is used to refer to the
273
transaction record when there are conflicts and to make
274
tie-breaking decisions on ordering between identical timestamps.
276
original candidate timestamp in the absence of read/write conflicts);
277
the client selects the maximum from amongst all write timestamps as the
278
final commit timestamp.
279
280
2. Commit the transaction by updating its transaction record. The value
281
of the commit entry contains the candidate timestamp (increased as
282
necessary to accommodate any latest read timestamps). Note that the
283
transaction is considered fully committed at this point and control
284
may be returned to the client.
285
286
In the case of an SI transaction, a commit timestamp which was
287
increased to accommodate concurrent readers is perfectly
288
acceptable and the commit may continue. For SSI transactions,
289
however, a gap between candidate and commit timestamps
290
necessitates transaction restart (note: restart is different than
291
abort--see below).
292
293
After the transaction is committed, all written intents are upgraded
294
in parallel by removing the “intent” flag. The transaction is
295
considered fully committed before this step and does not wait for
296
it to return control to the transaction coordinator.
297
298
In the absence of conflicts, this is the end. Nothing else is necessary
299
to ensure the correctness of the system.
300
301
**Conflict Resolution**
302
303
Things get more interesting when a reader or writer encounters an intent
304
record or newly-committed value in a location that it needs to read or
305
write. This is a conflict, usually causing either of the transactions to
306
abort or restart depending on the type of conflict.
307
308
***Transaction restart:***
309
310
This is the usual (and more efficient) type of behaviour and is used
311
except when the transaction was aborted (for instance by another
312
transaction).
313
In effect, that reduces to two cases; the first being the one outlined
314
above: An SSI transaction that finds upon attempting to commit that
315
its commit timestamp has been pushed. The second case involves a transaction
316
actively encountering a conflict, that is, one of its readers or writers
317
encounter data that necessitate conflict resolution
318
(see transaction interactions below).
319
320
When a transaction restarts, it changes its priority and/or moves its
323
have written some write intents, which need to be deleted before the
324
transaction commits, so as to not be included as part of the transaction.
325
These stale write intent deletions are done during the reexecution of the
327
the same keys as part of the reexecution of the transaction, or explicitly,
328
by cleaning up stale intents that are not part of the reexecution of the
329
transaction. Since most transactions will end up writing to the same keys,
330
the explicit cleanup run just before committing the transaction is usually
331
a NOOP.
332
333
***Transaction abort:***
334
335
This is the case in which a transaction, upon reading its transaction
336
record, finds that it has been aborted. In this case, the transaction
337
can not reuse its intents; it returns control to the client before
338
cleaning them up (other readers and writers would clean up dangling
339
intents as they encounter them) but will make an effort to clean up
340
after itself. The next attempt (if applicable) then runs as a new
341
transaction with **a new txn id**.
344
345
There are several scenarios in which transactions interact:
346
347
- **Reader encounters write intent or value with newer timestamp far
348
enough in the future**: This is not a conflict. The reader is free
349
to proceed; after all, it will be reading an older version of the
350
value and so does not conflict. Recall that the write intent may
351
be committed with a later timestamp than its candidate; it will
353
354
- **Reader encounters write intent or value with newer timestamp in the
355
near future:** In this case, we have to be careful. The newer
356
intent may, in absolute terms, have happened in our read's past if
357
the clock of the writer is ahead of the node serving the values.
358
In that case, we would need to take this value into account, but
359
we just don't know. Hence the transaction restarts, using instead
360
a future timestamp (but remembering a maximum timestamp used to
362
this is optimized further; see the details under "choosing a time
363
stamp" below.
364
365
- **Reader encounters write intent with older timestamp**: the reader
367
If the transaction has already been committed, then the reader can
368
just read the value. If the write transaction has not yet been
369
committed, then the reader has two options. If the write conflict
370
is from an SI transaction, the reader can *push that transaction's
371
commit timestamp into the future* (and consequently not have to
372
read it). This is simple to do: the reader just updates the
373
transaction’s commit timestamp to indicate that when/if the
374
transaction does commit, it should use a timestamp *at least* as
375
high. However, if the write conflict is from an SSI transaction,
376
the reader must compare priorities. If the reader has the higher priority,
377
it pushes the transaction’s commit timestamp (that
378
transaction will then notice its timestamp has been pushed, and
380
a new priority `max(new random priority, conflicting txn’s
382
385
priority, the writer aborts the conflicting transaction. If the write
386
intent has a higher or equal priority the transaction retries, using as a new
387
priority *max(new random priority, conflicting txn’s priority - 1)*;
388
the retry occurs after a short, randomized backoff interval.
389
390
- **Writer encounters newer committed value**:
391
The committed value could also be an unresolved write intent made by a
392
transaction that has already committed. The transaction restarts. On restart,
393
the same priority is reused, but the candidate timestamp is moved forward
394
to the encountered value's timestamp.
395
398
candidate timestamp is earlier than the low water mark on the cache itself
399
(i.e. its last evicted timestamp) or if the key being written has a read
400
timestamp later than the write’s candidate timestamp, this later timestamp
401
value is returned with the write. A new timestamp forces a transaction
402
restart only if it is serializable.
404
**Transaction management**
405
406
Transactions are managed by the client proxy (or gateway in SQL Azure
407
parlance). Unlike in Spanner, writes are not buffered but are sent
408
directly to all implicated ranges. This allows the transaction to abort
409
quickly if it encounters a write conflict. The client proxy keeps track
410
of all written keys in order to resolve write intents asynchronously upon
411
transaction completion. If a transaction commits successfully, all intents
412
are upgraded to committed. In the event a transaction is aborted, all written
413
intents are deleted. The client proxy doesn’t guarantee it will resolve intents.
414
415
In the event the client proxy restarts before the pending transaction is
416
committed, the dangling transaction would continue to "live" until
417
aborted by another transaction. Transactions periodically heartbeat
418
their transaction record to maintain liveness.
419
Transactions encountered by readers or writers with dangling intents
420
which haven’t been heartbeat within the required interval are aborted.
425
426
An exploration of retries with contention and abort times with abandoned
427
transaction is
428
[here](https://docs.google.com/document/d/1kBCu4sdGAnvLqpT-_2vaTbomNmX3_saayWEGYu1j7mQ/edit?usp=sharing).
429
431
432
Please see [pkg/roachpb/data.proto](https://github.com/cockroachdb/cockroach/blob/master/pkg/roachpb/data.proto) for the up-to-date structures, the best entry point being `message Transaction`.
433
434
**Pros**
435
436
- No requirement for reliable code execution to prevent stalled 2PC
437
protocol.
438
- Readers never block with SI semantics; with SSI semantics, they may
439
abort.
440
- Lower latency than traditional 2PC commit protocol (w/o contention)
441
because second phase requires only a single write to the
443
transaction participants.
444
- Priorities avoid starvation for arbitrarily long transactions and
445
always pick a winner from between contending transactions (no
446
mutual aborts).
447
- Writes not buffered at client; writes fail fast.
448
- No read-locking overhead required for *serializable* SI (in contrast
449
to other SSI implementations).
450
- Well-chosen (i.e. less random) priorities can flexibly give
451
probabilistic guarantees on latency for arbitrary transactions
452
(for example: make OLTP transactions 10x less likely to abort than
453
low priority transactions, such as asynchronously scheduled jobs).
454
455
**Cons**
456
459
- Abandoned transactions may block contending writers for up to the
460
heartbeat interval, though average wait is likely to be
461
considerably shorter (see [graph in link](https://docs.google.com/document/d/1kBCu4sdGAnvLqpT-_2vaTbomNmX3_saayWEGYu1j7mQ/edit?usp=sharing)).
462
This is likely considerably more performant than detecting and
463
restarting 2PC in order to release read and write locks.
464
- Behavior different than other SI implementations: no first writer
465
wins, and shorter transactions do not always finish quickly.
466
Element of surprise for OLTP systems may be a problematic factor.
467
- Aborts can decrease throughput in a contended system compared with
468
two phase locking. Aborts and retries increase read and write
469
traffic, increase latency and decrease throughput.
470
471
**Choosing a Timestamp**
472
474
is choosing a timestamp guaranteed to be greater than the latest
475
timestamp of any committed transaction (in absolute time). No system can
476
claim consistency and fail to read already-committed data.
477
479
accessing a single node is easy. The timestamp is assigned by the node
480
itself, so it is guaranteed to be at a greater timestamp than all the
481
existing timestamped data on the node.
482
483
For multiple nodes, the timestamp of the node coordinating the
484
transaction `t` is used. In addition, a maximum timestamp `t+ε` is
485
supplied to provide an upper bound on timestamps for already-committed
487
data read which have timestamps greater than `t` but less than `t+ε`
488
cause the transaction to abort and retry with the conflicting timestamp
490
the same. This implies that transaction restarts due to clock uncertainty
491
can only happen on a time interval of length `ε`.
495
into account t<sub>c</sub>, but the timestamp of the node at the time
496
of the uncertain read t<sub>node</sub>. The larger of those two timestamps
497
t<sub>c</sub> and t<sub>node</sub> (likely equal to the latter) is used
498
to increase the read timestamp. Additionally, the conflicting node is
499
marked as “certain”. Then, for future reads to that node within the
500
transaction, we set `MaxTimestamp = Read Timestamp`, preventing further
501
uncertainty restarts.
502
503
Correctness follows from the fact that we know that at the time of the read,
504
there exists no version of any key on that node with a higher timestamp than
506
encounters a key with a higher timestamp, it knows that in absolute time,
507
the value was written after t<sub>node</sub> was obtained, i.e. after the
508
uncertain read. Hence the transaction can move forward reading an older version
509
of the data (at the transaction's timestamp). This limits the time uncertainty
510
restarts attributed to a node to at most one. The tradeoff is that we might
511
pick a timestamp larger than the optimal one (> highest conflicting timestamp),
512
resulting in the possibility of a few more conflicts.
513
514
We expect retries will be rare, but this assumption may need to be
515
revisited if retries become problematic. Note that this problem does not
516
apply to historical reads. An alternate approach which does not require
517
retries makes a round to all node participants in advance and
518
chooses the highest reported node wall time as the timestamp. However,
519
knowing which nodes will be accessed in advance is difficult and
520
potentially limiting. Cockroach could also potentially use a global
521
clock (Google did this with [Percolator](https://www.usenix.org/legacy/event/osdi10/tech/full_papers/Peng.pdf)),
522
which would be feasible for smaller, geographically-proximate clusters.
523
524
# Strict Serializability (Linearizability)
525
526
Roughly speaking, the gap between <i>strict serializability</i> (which we use
527
interchangeably with <i>linearizability</i>) and CockroachDB's default
528
isolation level (<i>serializable</i>) is that with linearizable transactions,
529
causality is preserved. That is, if one transaction (say, creating a posting
530
for a user) waits for its predecessor (creating the user in the first place)
531
to complete, one would hope that the logical timestamp assigned to the former
532
is larger than that of the latter.
533
In practice, in distributed databases this may not hold, the reason typically
534
being that clocks across a distributed system are not perfectly synchronized
535
and the "later" transaction touches a part disjoint from that on which the
536
first transaction ran, resulting in clocks with disjoint information to decide
537
on the commit timestamps.
538
539
In practice, in CockroachDB many transactional workloads are actually
540
linearizable, though the precise conditions are too involved to outline them
541
here.
542
543
Causality is typically not required for many transactions, and so it is
544
advantageous to pay for it only when it *is* needed. CockroachDB implements
545
this via <i>causality tokens</i>: When committing a transaction, a causality
546
token can be retrieved and passed to the next transaction, ensuring that these
547
two transactions get assigned increasing logical timestamps.
548
549
Additionally, as better synchronized clocks become a standard commodity offered
550
by cloud providers, CockroachDB can provide global linearizability by doing
551
much the same that [Google's
552
Spanner](http://research.google.com/archive/spanner.html) does: wait out the
553
maximum clock offset after committing, but before returning to the client.
554
555
See the blog post below for much more in-depth information.
556
557
https://www.cockroachlabs.com/blog/living-without-atomic-clocks/
558
559
# Logical Map Content
560
561
Logically, the map contains a series of reserved system key/value
562
pairs preceding the actual user data (which is managed by the SQL
563
subsystem).
564
566
- ...
567
- `\x02<keyN>`: Range metadata for range ending `\x03<keyN>`. This a "meta1" key.
568
- `\x03<key1>`: Range metadata for range ending `<key1>`. This a "meta2" key.
569
- ...
570
- `\x03<keyN>`: Range metadata for range ending `<keyN>`. This a "meta2" key.
571
- `\x04{desc,node,range,store}-idegen`: ID generation oracles for various component types.
572
- `\x04status-node-<varint encoded Store ID>`: Store runtime metadata.
573
- `\x04tsd<key>`: Time-series data key.
574
- `<key>`: A user key. In practice, these keys are managed by the SQL
575
subsystem, which employs its own key anatomy.
576
577
# Stores and Storage
578
579
Nodes contain one or more stores. Each store should be placed on a unique disk.
580
Internally, each store contains a single instance of RocksDB with a block cache
581
shared amongst all of the stores in a node. And these stores in turn have
582
a collection of range replicas. More than one replica for a range will never
583
be placed on the same store or even the same node.
584
585
Early on, when a cluster is first initialized, the few default starting ranges
586
will only have a single replica, but as soon as other nodes are available they
587
will replicate to them until they've reached their desired replication factor,
588
the default being 3.
589
590
Zone configs can be used to control a range's replication factor and add
591
constraints as to where the range's replicas can be located. When there is a
592
change in a range's zone config, the range will up or down replicate to the
593
appropriate number of replicas and move its replicas to the appropriate stores
594
based on zone config's constraints.
595
596
# Self Repair
597
598
If a store has not been heard from (gossiped their descriptors) in some time,
599
the default setting being 5 minutes, the cluster will consider this store to be
600
dead. When this happens, all ranges that have replicas on that store are
601
determined to be unavailable and removed. These ranges will then upreplicate
602
themselves to other available stores until their desired replication factor is
603
again met. If 50% or more of the replicas are unavailable at the same time,
604
there is no quorum and the whole range will be considered unavailable until at
605
least greater than 50% of the replicas are again available.
606
607
# Rebalancing
608
609
As more data are added to the system, some stores may grow faster than others.
610
To combat this and to spread the overall load across the full cluster, replicas
611
will be moved between stores maintaining the desired replication factor. The
612
heuristics used to perform this rebalancing include:
613
614
- the number of replicas per store
615
- the total size of the data used per store
616
- free space available per store
617
618
In the future, some other factors that might be considered include:
619
620
- cpu/network load per store
621
- ranges that are used together often in queries
622
- number of active ranges per store
623
- number of range leases held per store
624
625
# Range Metadata
626
627
The default approximate size of a range is 64M (2\^26 B). In order to
628
support 1P (2\^50 B) of logical data, metadata is needed for roughly
629
2\^(50 - 26) = 2\^24 ranges. A reasonable upper bound on range metadata
632
B would require roughly 4G (2\^32 B) to store--too much to duplicate
633
between machines. Our conclusion is that range metadata must be
634
distributed for large installations.
635
636
To keep key lookups relatively fast in the presence of distributed metadata,
637
we store all the top-level metadata in a single range (the first range). These
638
top-level metadata keys are known as *meta1* keys, and are prefixed such that
639
they sort to the beginning of the key space. Given the metadata size of 256
640
bytes given above, a single 64M range would support 64M/256B = 2\^18 ranges,
642
above, we need two levels of indirection, where the first level addresses the
643
second, and the second addresses user data. With two levels of indirection, we
644
can address 2\^(18 + 18) = 2\^36 ranges; each range addresses 2\^26 B, and
645
altogether we address 2\^(36+26) B = 2\^62 B = 4E of user data.
646
647
For a given user-addressable `key1`, the associated *meta1* record is found
648
at the successor key to `key1` in the *meta1* space. Since the *meta1* space
649
is sparse, the successor key is defined as the next key which is present. The
650
*meta1* record identifies the range containing the *meta2* record, which is
651
found using the same process. The *meta2* record identifies the range
652
containing `key1`, which is again found the same way (see examples below).
653
654
Concretely, metadata keys are prefixed by `\x02` (meta1) and `\x03`
655
(meta2); the prefixes `\x02` and `\x03` provide for the desired
656
sorting behaviour. Thus, `key1`'s *meta1* record will reside at the
657
successor key to `\x02<key1>`.
660
the RocksDB iterator only supports a Seek() interface which acts as a
661
Ceil(). Using the start key of the range would cause Seek() to find the
662
key *after* the meta indexing record we’re looking for, which would
663
result in having to back the iterator up, an option which is both less
664
efficient and not available in all cases.
665
666
The following example shows the directory structure for a map with
667
three ranges worth of data. Ellipses indicate additional key/value
668
pairs to fill an entire range of data. For clarity, the examples use
669
`meta1` and `meta2` to refer to the prefixes `\x02` and `\x03`. Except
670
for the fact that splitting ranges requires updates to the range
671
metadata with knowledge of the metadata layout, the range metadata
672
itself requires no special treatment or bootstrapping.
673
674
**Range 0** (located on servers `dcrama1:8000`, `dcrama2:8000`,
675
`dcrama3:8000`)
676
677
- `meta1\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
678
- `meta2<lastkey0>`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
679
- `meta2<lastkey1>`: `dcrama4:8000`, `dcrama5:8000`, `dcrama6:8000`
680
- `meta2\xff`: `dcrama7:8000`, `dcrama8:8000`, `dcrama9:8000`
681
- ...
682
- `<lastkey0>`: `<lastvalue0>`
683
684
**Range 1** (located on servers `dcrama4:8000`, `dcrama5:8000`,
685
`dcrama6:8000`)
686
687
- ...
688
- `<lastkey1>`: `<lastvalue1>`
689
690
**Range 2** (located on servers `dcrama7:8000`, `dcrama8:8000`,
691
`dcrama9:8000`)
692
693
- ...
694
- `<lastkey2>`: `<lastvalue2>`
695
696
Consider a simpler example of a map containing less than a single
697
range of data. In this case, all range metadata and all data are
698
located in the same range:
699
700
**Range 0** (located on servers `dcrama1:8000`, `dcrama2:8000`,
701
`dcrama3:8000`)*
702
703
- `meta1\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
704
- `meta2\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
705
- `<key0>`: `<value0>`
706
- `...`
707
708
Finally, a map large enough to need both levels of indirection would
709
look like (note that instead of showing range replicas, this
710
example is simplified to just show range indexes):
711
712
**Range 0**
713
714
- `meta1<lastkeyN-1>`: Range 0
715
- `meta1\xff`: Range 1
716
- `meta2<lastkey1>`: Range 1
717
- `meta2<lastkey2>`: Range 2
718
- `meta2<lastkey3>`: Range 3
719
- ...
721
722
**Range 1**
723
726
- ...
728
- ...
729
- `<lastkey1>`: `<lastvalue1>`
730
731
**Range 2**
732
733
- ...
734
- `<lastkey2>`: `<lastvalue2>`
735
736
**Range 3**
737
738
- ...
739
- `<lastkey3>`: `<lastvalue3>`
740
741
**Range 262144**
742
743
- ...
744
- `<lastkeyN>`: `<lastvalueN>`
745
746
**Range 262145**
747
748
- ...
749
- `<lastkeyN+1>`: `<lastvalueN+1>`
750
751
Note that the choice of range `262144` is just an approximation. The
752
actual number of ranges addressable via a single metadata range is
753
dependent on the size of the keys. If efforts are made to keep key sizes
754
small, the total number of addressable ranges would increase and vice
755
versa.
756
757
From the examples above it’s clear that key location lookups require at
758
most three reads to get the value for `<key>`:
759
762
3. `<key>`.
763
764
For small maps, the entire lookup is satisfied in a single RPC to Range 0. Maps
765
containing less than 16T of data would require two lookups. Clients cache both
766
levels of range metadata, and we expect that data locality for individual
767
clients will be high. Clients may end up with stale cache entries. If on a
768
lookup, the range consulted does not match the client’s expectations, the
769
client evicts the stale entries and possibly does a new lookup.
770
773
Each range is configured to consist of three or more replicas, as specified by
774
their ZoneConfig. The replicas in a range maintain their own instance of a
775
distributed consensus algorithm. We use the [*Raft consensus algorithm*](https://raftconsensus.github.io)
778
[ePaxos](https://www.cs.cmu.edu/~dga/papers/epaxos-sosp2013.pdf) has
779
promising performance characteristics for WAN-distributed replicas, but
780
it does not guarantee a consistent ordering between replicas.
781
782
Raft elects a relatively long-lived leader which must be involved to
784
replicated. In the absence of heartbeats, followers become candidates
785
after randomized election timeouts and proceed to hold new leader
786
elections. Cockroach weights random timeouts such that the replicas with
787
shorter round trip times to peers are more likely to hold elections
788
first (not implemented yet). Only the Raft leader may propose commands;
789
followers will simply relay commands to the last known leader.
791
Our Raft implementation was developed together with CoreOS, but adds an extra
792
layer of optimization to account for the fact that a single Node may have
793
millions of consensus groups (one for each Range). Areas of optimization
794
are chiefly coalesced heartbeats (so that the number of nodes dictates the
795
number of heartbeats as opposed to the much larger number of ranges) and
796
batch processing of requests.
797
Future optimizations may include two-phase elections and quiescent ranges
798
(i.e. stopping traffic completely for inactive ranges).
799
801
802
As outlined in the Raft section, the replicas of a Range are organized as a
803
Raft group and execute commands from their shared commit log. Going through
804
Raft is an expensive operation though, and there are tasks which should only be
805
carried out by a single replica at a time (as opposed to all of them).
806
In particular, it is desirable to serve authoritative reads from a single
807
Replica (ideally from more than one, but that is far more difficult).
810
This is a lease held for a slice of (database, i.e. hybrid logical) time.
811
A replica establishes itself as owning the lease on a range by committing
812
a special lease acquisition log entry through raft. The log entry contains
813
the replica node's epoch from the node liveness table--a system
814
table containing an epoch and an expiration time for each node. A node is
815
responsible for continuously updating the expiration time for its entry
816
in the liveness table. Once the lease has been committed through raft
817
the replica becomes the lease holder as soon as it applies the lease
818
acquisition command, guaranteeing that when it uses the lease it has
819
already applied all prior writes on the replica and can see them locally.
820
821
To prevent two nodes from acquiring the lease, the requestor includes a copy
822
of the lease that it believes to be valid at the time it requests the lease.
823
If that lease is still valid when the new lease is applied, it is granted,
824
or another lease is granted in the interim and the requested lease is
825
ignored. A lease can move from node A to node B only after node A's
826
liveness record has expired and its epoch has been incremented.
827
828
Note: range leases for ranges within the node liveness table keyspace and
829
all ranges that precede it, including meta1 and meta2, are not managed using
830
the above mechanism to prevent circular dependencies.
831
832
A replica holding a lease at a specific epoch can use the lease as long as
833
the node epoch hasn't changed and the expiration time hasn't passed.
834
The replica holding the lease may satisfy reads locally, without incurring the
835
overhead of going through Raft, and is in charge or involved in handling
836
Range-specific maintenance tasks such as splitting, merging and rebalancing
837
838
All Reads and writes are generally addressed to the replica holding
839
the lease; if none does, any replica may be addressed, causing it to try
840
to obtain the lease synchronously. Requests received by a non-lease holder
841
(for the HLC timestamp specified in the request's header) fail with an
842
error pointing at the replica's last known lease holder. These requests
843
are retried transparently with the updated lease by the gateway node and
844
never reach the client.
845
846
Since reads bypass Raft, a new lease holder will, among other things, ascertain
847
that its timestamp cache does not report timestamps smaller than the previous
848
lease holder's (so that it's compatible with reads which may have occurred on
849
the former lease holder). This is accomplished by letting leases enter
850
a <i>stasis period</i> (which is just the expiration minus the maximum clock
851
offset) before the actual expiration of the lease, so that all the next lease
852
holder has to do is set the low water mark of the timestamp cache to its
853
new lease's start time.
854
855
As a lease enters its stasis period, no more reads or writes are served, which
856
is undesirable. However, this would only happen in practice if a node became
857
unavailable. In almost all practical situations, no unavailability results
858
since leases are usually long-lived (and/or eagerly extended, which can avoid
859
the stasis period) or proactively transferred away from the lease holder, which
860
can also avoid the stasis period by promising not to serve any further reads
861
until the next lease goes into effect.
862
863
## Colocation with Raft leadership
866
further efforts, Raft leadership and the Range lease might not be held by the
867
same Replica. Since it's expensive to not have these two roles colocated (the
868
lease holder has to forward each proposal to the leader, adding costly RPC
869
round-trips), each lease renewal or transfer also attempts to colocate them.
870
In practice, that means that the mismatch is rare and self-corrects quickly.
874
This subsection describes how a lease holder replica processes a
875
read/write command in more details. Each command specifies (1) a key
876
(or a range of keys) that the command accesses and (2) the ID of a
877
range which the key(s) belongs to. When receiving a command, a node
878
looks up a range by the specified Range ID and checks if the range is
879
still responsible for the supplied keys. If any of the keys do not
880
belong to the range, the node returns an error so that the client will
881
retry and send a request to a correct range.
884
process the command. If the command is an inconsistent read-only
885
command, it is processed immediately. If the command is a consistent
886
read or a write, the command is executed when both of the following
887
conditions hold:
888
890
- There are no other running commands whose keys overlap with
891
the submitted command and cause read/write conflict.
892
893
When the first condition is not met, the replica attempts to acquire
894
a lease or returns an error so that the client will redirect the
899
When the above two conditions are met, the lease holder replica processes the
900
command. Consistent reads are processed on the lease holder immediately.
902
will execute the same commands. All commands produce deterministic
903
results so that the range replicas keep consistent states among them.
904
905
When a write command completes, all the replica updates their response
907
replica updates its timestamp cache to keep track of the latest read
908
for a given key.
909
911
executed. Before executing a command, each replica checks if a replica
912
proposing the command has a still lease. When the lease has been
913
expired, the command will be rejected by the replica.
914
915
919
minimum thresholds for capacity or load. Ranges exceeding maximums for
920
either capacity or load are split; ranges below minimums for *both*
921
capacity and load are merged.
922
923
Ranges maintain the same accounting statistics as accounting key
924
prefixes. These boil down to a time series of data points with minute
925
granularity. Everything from number of bytes to read/write queue sizes.
926
Arbitrary distillations of the accounting stats can be determined as the
928
split/merge are range size in bytes and IOps. A good metric for
929
rebalancing a replica from one node to another would be total read/write
930
queue wait times. These metrics are gossipped, with each range / node
931
passing along relevant metrics if they’re in the bottom or top of the
932
range it’s aware of.
933
934
A range finding itself exceeding either capacity or load threshold
936
candidate and issues the split through Raft. In contrast to splitting,
937
merging requires a range to be below the minimum threshold for both
938
capacity *and* load. A range being merged chooses the smaller of the
939
ranges immediately preceding and succeeding it.
940
941
Splitting, merging, rebalancing and recovering all follow the same basic
942
algorithm for moving data between roach nodes. New target replicas are
943
created and added to the replica set of source range. Then each new
944
replica is brought up to date by either replaying the log in full or
945
copying a snapshot of the source replica data and then replaying the log
946
from the timestamp of the snapshot to catch up fully. Once the new
947
replicas are fully up to date, the range metadata is updated and old,
948
source replica(s) deleted if applicable.
949
951
952
```
953
if splitting
954
SplitRange(split_key): splits happen locally on range replicas and
955
only after being completed locally, are moved to new target replicas.
956
else if merging
957
Choose new replicas on same servers as target range replicas;
958
add to replica set.
959
else if rebalancing || recovering
960
Choose new replica(s) on least loaded servers; add to replica set.
961
```
962
963
**New Replica**
964
965
*Bring replica up to date:*
966
967
```
968
if all info can be read from replicated log
969
copy replicated log
970
else
971
snapshot source replica
972
send successive ReadRange requests to source replica
973
referencing snapshot
974
975
if merging
976
combine ranges on all replicas
977
else if rebalancing || recovering
978
remove old range replica(s)
979
```
980
982
configurable maximum threshold. Similarly, ranges are merged when the
983
total data falls below a configurable minimum threshold.
984
985
**TBD: flesh this out**: Especially for merges (but also rebalancing) we have a
986
range disappearing from the local node; that range needs to disappear
987
gracefully, with a smooth handoff of operation to the new owner of its data.
988
989
Ranges are rebalanced if a node determines its load or capacity is one
990
of the worst in the cluster based on gossipped load stats. A node with
991
spare capacity is chosen in the same datacenter and a special-case split
992
is done which simply duplicates the data 1:1 and resets the range
993
configuration metadata.
994
995
# Node Allocation (via Gossip)
996
997
New nodes must be allocated when a range is split. Instead of requiring
999
of peer nodes --or-- alternatively requiring a specialized curator or
1000
master with sufficiently global knowledge, we use a gossip protocol to
1001
efficiently communicate only interesting information between all of the
1002
nodes in the cluster. What’s interesting information? One example would
1003
be whether a particular node has a lot of spare capacity. Each node,
1004
when gossiping, compares each topic of gossip to its own state. If its
1005
own state is somehow “more interesting” than the least interesting item
1006
in the topic it’s seen recently, it includes its own state as part of
1007
the next gossip session with a peer node. In this way, a node with
1008
capacity sufficiently in excess of the mean quickly becomes discovered
1009
by the entire cluster. To avoid piling onto outliers, nodes from the
1010
high capacity set are selected at random for allocation.
1011
1012
The gossip protocol itself contains two primary components:
1013
1014
- **Peer Selection**: each node maintains up to N peers with which it
1015
regularly communicates. It selects peers with an eye towards
1016
maximizing fanout. A peer node which itself communicates with an
1017
array of otherwise unknown nodes will be selected over one which
1018
communicates with a set containing significant overlap. Each time
1019
gossip is initiated, each nodes’ set of peers is exchanged. Each
1020
node is then free to incorporate the other’s peers as it sees fit.
1021
To avoid any node suffering from excess incoming requests, a node
1022
may refuse to answer a gossip exchange. Each node is biased
1023
towards answering requests from nodes without significant overlap
1024
and refusing requests otherwise.
1025
1026
Peers are efficiently selected using a heuristic as described in
1027
[Agarwal & Trachtenberg (2006)](https://drive.google.com/file/d/0B9GCVTp_FHJISmFRTThkOEZSM1U/edit?usp=sharing).
1028
1029
**TBD**: how to avoid partitions? Need to work out a simulation of
1030
the protocol to tune the behavior and see empirically how well it
1031
works.
1032
1033
- **Gossip Selection**: what to communicate. Gossip is divided into
1034
topics. Load characteristics (capacity per disk, cpu load, and
1035
state [e.g. draining, ok, failure]) are used to drive node
1036
allocation. Range statistics (range read/write load, missing
1037
replicas, unavailable ranges) and network topology (inter-rack
1038
bandwidth/latency, inter-datacenter bandwidth/latency, subnet
1039
outages) are used for determining when to split ranges, when to
1040
recover replicas vs. wait for network connectivity, and for
1041
debugging / sysops. In all cases, a set of minimums and a set of
1042
maximums is propagated; each node applies its own view of the
1043
world to augment the values. Each minimum and maximum value is
1044
tagged with the reporting node and other accompanying contextual
1045
information. Each topic of gossip has its own protobuf to hold the
1046
structured data. The number of items of gossip in each topic is
1047
limited by a configurable bound.
1048
1049
For efficiency, nodes assign each new item of gossip a sequence
1050
number and keep track of the highest sequence number each peer
1051
node has seen. Each round of gossip communicates only the delta
1052
containing new items.
1053
1054
# Node and Cluster Metrics
1055
1056
Every component of the system is responsible for exporting interesting
1057
metrics about itself. These could be histograms, throughput counters, or
1058
gauges.
1059
1060
These metrics are exported for external monitoring systems (such as Prometheus)
1061
via a HTTP endpoint, but CockroachDB also implements an internal timeseries
1062
database which is stored in the replicated key-value map.
1063
1064
Time series are stored at Store granularity and allow the admin dashboard
1065
to efficiently gain visibility into a universe of information at the Cluster,
1066
Node or Store level. A [periodic background process](RFCS/20160901_time_series_culling.md)
1068
1069
# Zones
1070
1071
Zones provide a method for configuring the replication of portions of the
1072
keyspace. Zone values specify a protobuf containing
1073
the datacenters from which replicas for ranges which fall under
1074
the zone must be chosen.
1075
1077
[pkg/config/zone.proto](https://github.com/cockroachdb/cockroach/blob/master/pkg/config/zone.proto)
1078
for up-to-date data structures used, the best entry point being
1079
`message ZoneConfig`.
1080
1082
existing zones for its ranges against the zone configuration. If
1083
it discovers differences, it reconfigures ranges in the same way
1084
that it rebalances away from busy nodes, via special-case 1:1
1085
split to a duplicate range comprising the new configuration.
1086
1087
# SQL
1088
1089
Each node in a cluster can accept SQL client connections. CockroachDB
1090
supports the PostgreSQL wire protocol, to enable reuse of native
1091
PostgreSQL client drivers. Connections using SSL and authenticated
1092
using client certificates are supported and even encouraged over
1093
unencrypted (insecure) and password-based connections.
1094
1095
Each connection is associated with a SQL session which holds the
1096
server-side state of the connection. Over the lifespan of a session
1097
the client can send SQL to open/close transactions, issue statements
1098
or queries or configure session parameters, much like with any other
1099
SQL database.
1100
1101
## Language support
1102
1103
CockroachDB also attempts to emulate the flavor of SQL supported by
1104
PostgreSQL, although it also diverges in significant ways:
1105
1106
- CockroachDB exclusively implements MVCC-based consistency for
1107
transactions, and thus only supports SQL's isolation levels SNAPSHOT
1108
and SERIALIZABLE. The other traditional SQL isolation levels are
1109
internally mapped to either SNAPSHOT or SERIALIZABLE.
1110
1112
which only supports a limited form of implicit coercions between
1113
types compared to PostgreSQL. The rationale is to keep the
1114
implementation simple and efficient, capitalizing on the observation
1115
that 1) most SQL code in clients is automatically generated with
1116
coherent typing already and 2) existing SQL code for other databases
1117
will need to be massaged for CockroachDB anyways.
1118
1119
## SQL architecture
1120
1121
Client connections over the network are handled in each node by a
1122
pgwire server process (goroutine). This handles the stream of incoming
1123
commands and sends back responses including query/statement results.
1124
The pgwire server also handles pgwire-level prepared statements,
1125
binding prepared statements to arguments and looking up prepared
1126
statements for execution.
1127
1128
Meanwhile the state of a SQL connection is maintained by a Session
1129
object and a monolithic `planner` object (one per connection) which
1130
coordinates execution between the session, the current SQL transaction
1131
state and the underlying KV store.
1132
1133
Upon receiving a query/statement (either directly or via an execute
1134
command for a previously prepared statement) the pgwire server forwards
1135
the SQL text to the `planner` associated with the connection. The SQL
1136
code is then transformed into a SQL query plan.
1137
The query plan is implemented as a tree of objects which describe the
1138
high-level data operations needed to resolve the query, for example
1139
"join", "index join", "scan", "group", etc.
1140
1141
The query plan objects currently also embed the run-time state needed
1142
for the execution of the query plan. Once the SQL query plan is ready,
1143
methods on these objects then carry the execution out in the fashion
1144
of "generators" in other programming languages: each node *starts* its
1145
children nodes and from that point forward each child node serves as a
1146
*generator* for a stream of result rows, which the parent node can
1147
consume and transform incrementally and present to its own parent node
1148
also as a generator.
1149
1150
The top-level planner consumes the data produced by the top node of
1151
the query plan and returns it to the client via pgwire.
1152
1153
## Data mapping between the SQL model and KV
1154
1155
Every SQL table has a primary key in CockroachDB. (If a table is created
1156
without one, an implicit primary key is provided automatically.)
1157
The table identifier, followed by the value of the primary key for
1158
each row, are encoded as the *prefix* of a key in the underlying KV
1159
store.
1160
1161
Each remaining column or *column family* in the table is then encoded
1162
as a value in the underlying KV store, and the column/family identifier
1163
is appended as *suffix* to the KV key.
1164
1165
For example:
1166
1167
- after table `customers` is created in a database `mydb` with a
1168
primary key column `name` and normal columns `address` and `URL`, the KV pairs
1169
to store the schema would be:
1170
1171
| Key | Values |
1172
| ---------------------------- | ------ |
1173
| `/system/databases/mydb/id` | 51 |
1174
| `/system/tables/customer/id` | 42 |
1175
| `/system/desc/51/42/address` | 69 |
1176
| `/system/desc/51/42/url` | 66 |
1177
1178
(The numeric values on the right are chosen arbitrarily for the
1179
example; the structure of the schema keys on the left is simplified
1180
for the example and subject to change.) Each database/table/column
1181
name is mapped to a spontaneously generated identifier, so as to
1182
simplify renames.
1183
1184
Then for a single row in this table:
1185
1186
| Key | Values |
1187
| ----------------- | -------------------------------- |
1188
| `/51/42/Apple/69` | `1 Infinite Loop, Cupertino, CA` |
1190
1191
Each key has the table prefix `/51/42` followed by the primary key
1192
prefix `/Apple` followed by the column/family suffix (`/66`,
1193
`/69`). The KV value is directly encoded from the SQL value.
1194
1195
Efficient storage for the keys is guaranteed by the underlying RocksDB engine
1196
by means of prefix compression.
1197
1198
Finally, for SQL indexes, the KV key is formed using the SQL value of the
1199
indexed columns, and the KV value is the KV key prefix of the rest of
1200
the indexed row.
1201
1202
## Distributed SQL
1203
1204
Dist-SQL is a new execution framework being developed as of Q3 2016 with the
1205
goal of distributing the processing of SQL queries.
1206
See the [Distributed SQL
1208
for a detailed design of the subsystem; this section will serve as a summary.
1209
1210
Distributing the processing is desirable for multiple reasons:
1211
- Remote-side filtering: when querying for a set of rows that match a filtering
1212
expression, instead of querying all the keys in certain ranges and processing
1213
the filters after receiving the data on the gateway node over the network,
1214
we'd like the filtering expression to be processed by the lease holder or
1215
remote node, saving on network traffic and related processing.
1216
- For statements like `UPDATE .. WHERE` and `DELETE .. WHERE` we want to
1217
perform the query and the updates on the node which has the data (as opposed
1218
to receiving results at the gateway over the network, and then performing the
1219
update or deletion there, which involves additional round-trips).
1220
- Parallelize SQL computation: when significant computation is required, we
1221
want to distribute it to multiple node, so that it scales with the amount of
1222
data involved. This applies to `JOIN`s, aggregation, sorting.
1223
1224
The approach we took was originally inspired by
1225
[Sawzall](https://cloud.google.com/dataflow/model/programming-model) - a
1226
project by Rob Pike et al. at Google that proposes a "shell" (high-level
1227
language interpreter) to ease the exploitation of MapReduce. It provides a
1228
clear separation between "local" processes which process a limited amount of
1229
data and distributed computations, which are abstracted away behind a
1230
restricted set of conceptual constructs.
1231
1232
To run SQL statements in a distributed fashion, we introduce a couple of concepts:
1233
- _logical plan_ - similar on the surface to the `planNode` tree described in
1234
the [SQL](#sql) section, it represents the abstract (non-distributed) data flow
1235
through computation stages.
1236
- _physical plan_ - a physical plan is conceptually a mapping of the _logical
1237
plan_ nodes to CockroachDB nodes. Logical plan nodes are replicated and
1238
specialized depending on the cluster topology. The components of the physical
1239
plan are scheduled and run on the cluster.
1240
1241
## Logical planning
1242
1243
The logical plan is made up of _aggregators_. Each _aggregator_ consumes an
1244
_input stream_ of rows (or multiple streams for joins) and produces an _output
1245
stream_ of rows. Both the input and the output streams have a set schema. The
1246
streams are a logical concept and might not map to a single data stream in the
1247
actual computation. Aggregators will be potentially distributed when converting
1248
the *logical plan* to a *physical plan*; to express what distribution and
1249
parallelization is allowed, an aggregator defines a _grouping_ on the data that
1250
flows through it, expressing which rows need to be processed on the same node
1251
(this mechanism constraints rows matching in a subset of columns to be
1252
processed on the same node). This concept is useful for aggregators that need
1253
to see some set of rows for producing output - e.g. the SQL aggregation
1254
functions. An aggregator with no grouping is a special but important case in
1255
which we are not aggregating multiple pieces of data, but we may be filtering,
1256
transforming, or reordering individual pieces of data.
1257
1258
Special **table reader** aggregators with no inputs are used as data sources; a
1259
table reader can be configured to output only certain columns, as needed.
1260
A special **final** aggregator with no outputs is used for the results of the
1261
query/statement.
1262
1263
To reflect the result ordering that a query has to produce, some aggregators
1264
(`final`, `limit`) are configured with an **ordering requirement** on the input
1265
stream (a list of columns with corresponding ascending/descending
1266
requirements). Some aggregators (like `table readers`) can guarantee a certain
1267
ordering on their output stream, called an **ordering guarantee**. All
1268
aggregators have an associated **ordering characterization** function
1269
`ord(input_order) -> output_order` that maps `input_order` (an ordering
1270
guarantee on the input stream) into `output_order` (an ordering guarantee for
1271
the output stream) - meaning that if the rows in the input stream are ordered
1272
according to `input_order`, then the rows in the output stream will be ordered
1273
according to `output_order`.
1274
1275
The ordering guarantee of the table readers along with the characterization
1276
functions can be used to propagate ordering information across the logical plan.
1277
When there is a mismatch (an aggregator has an ordering requirement that is not
1278
matched by a guarantee), we insert a **sorting aggregator**.
1279
1280
### Types of aggregators
1281
1282
- `TABLE READER` is a special aggregator, with no input stream. It's configured
1283
with spans of a table or index and the schema that it needs to read.
1284
Like every other aggregator, it can be configured with a programmable output
1285
filter.
1286
- `JOIN` performs a join on two streams, with equality constraints between
1287
certain columns. The aggregator is grouped on the columns that are
1288
constrained to be equal.
1289
- `JOIN READER` performs point-lookups for rows with the keys indicated by the
1290
input stream. It can do so by performing (potentially remote) KV reads, or by
1291
setting up remote flows.
1292
- `SET OPERATION` takes several inputs and performs set arithmetic on them
1293
(union, difference).
1294
- `AGGREGATOR` is the one that does "aggregation" in the SQL sense. It groups
1295
rows and computes an aggregate for each group. The group is configured using
1296
the group key. `AGGREGATOR` can be configured with one or more aggregation
1297
functions:
1298
- `SUM`
1299
- `COUNT`
1300
- `COUNT DISTINCT`
1301
- `DISTINCT`
1302
1303
An optional output filter has access to the group key and all the
1304
aggregated values (i.e. it can use even values that are not ultimately
1305
outputted).
1307
This is a no-grouping aggregator, hence it can be distributed arbitrarily to
1308
the data producers. This means that it doesn't produce a global ordering,
1309
instead it just guarantees an intra-stream ordering on each physical output
1310
streams). The global ordering, when needed, is achieved by an input
1311
synchronizer of a grouped processor (such as `LIMIT` or `FINAL`).
1312
- `LIMIT` is a single-group aggregator that stops after reading so many input
1313
rows.
1314
- `FINAL` is a single-group aggregator, scheduled on the gateway, that collects
1315
the results of the query. This aggregator will be hooked up to the pgwire
1316
connection to the client.
1317
1318
## Physical planning
1319
1320
Logical plans are transformed into physical plans in a *physical planning
1321
phase*. See the [corresponding
1322
section](RFCS/20160421_distributed_sql.md#from-logical-to-physical) of the Distributed SQL RFC
1323
for details. To summarize, each aggregator is planned as one or more
1324
*processors*, which we distribute starting from the data layout - `TABLE
1325
READER`s have multiple instances, split according to the ranges - each instance
1326
is planned on the lease holder of the relevant range. From that point on,
1327
subsequent processors are generally either colocated with their inputs, or
1328
planned as singletons, usually on the final destination node.
1329
1330
### Processors
1331
1332
When turning a _logical plan_ into a _physical plan_, its nodes are turned into
1333
_processors_. Processors are generally made up of three components:
1334
1335
1336
1337
1. The *input synchronizer* merges the input streams into a single stream of
1338
data. Types:
1339
* single-input (pass-through)
1340
* unsynchronized: passes rows from all input streams, arbitrarily
1341
interleaved.
1342
* ordered: the input physical streams have an ordering guarantee (namely the
1343
guarantee of the corresponding logical stream); the synchronizer is careful
1344
to interleave the streams so that the merged stream has the same guarantee.
1345
1346
2. The *data processor* core implements the data transformation or aggregation
1347
logic (and in some cases performs KV operations).
1348
1349
3. The *output router* splits the data processor's output to multiple streams;
1350
types:
1351
* single-output (pass-through)
1352
* mirror: every row is sent to all output streams
1353
* hashing: each row goes to a single output stream, chosen according
1354
to a hash function applied on certain elements of the data tuples.
1355
* by range: the router is configured with range information (relating to a
1356
certain table) and is able to send rows to the nodes that are lease holders for
1357
the respective ranges (useful for `JoinReader` nodes (taking index values
1358
to the node responsible for the PK) and `INSERT` (taking new rows to their
1359
lease holder-to-be)).
1360
1361
To illustrate with an example from the Distributed SQL RFC, the query:
1362
```
1363
TABLE Orders (OId INT PRIMARY KEY, CId INT, Value DECIMAL, Date DATE)
1364
1365
SELECT CID, SUM(VALUE) FROM Orders
1366
WHERE DATE > 2015
1367
GROUP BY CID
1368
ORDER BY 1 - SUM(Value)
1369
```
1370
1371
produces the following logical plan:
1372
1373
1374
1375
This logical plan above could be transformed into either one of the following
1376
physical plans:
1377
1378
1379
1380
or
1381
1382
1383
1384
1385
## Execution infrastructure
1386
1387
Once a physical plan has been generated, the system needs to divvy it up
1388
between the nodes and send it around for execution. Each node is responsible
1389
for locally scheduling data processors and input synchronizers. Nodes also
1390
communicate with each other for connecting output routers to input
1391
synchronizers through a streaming interface.
1392
1393
### Creating a local plan: the `ScheduleFlows` RPC
1394
1395
Distributed execution starts with the gateway making a request to every node
1396
that's supposed to execute part of the plan asking the node to schedule the
1397
sub-plan(s) it's responsible for (except for "on-the-fly" flows, see design
1398
doc). A node might be responsible for multiple disparate pieces of the overall
1399
DAG - let's call each of them a *flow*. A flow is described by the sequence of
1400
physical plan nodes in it, the connections between them (input synchronizers,
1401
output routers) plus identifiers for the input streams of the top node in the
1402
plan and the output streams of the (possibly multiple) bottom nodes. A node
1403
might be responsible for multiple heterogeneous flows. More commonly, when a
1404
node is the lease holder for multiple ranges from the same table involved in
1405
the query, it will run a `TableReader` configured with all the spans to be
1406
read across all the ranges local to the node.
1407
1408
A node therefore implements a `ScheduleFlows` RPC which takes a set of flows,
1409
sets up the input and output [mailboxes](#mailboxes), creates the local
1410
processors and starts their execution.
1411
1412
### Local scheduling of flows
1413
1414
The simplest way to schedule the different processors locally on a node is
1415
concurrently: each data processor, synchronizer and router runs as a goroutine,
1416
with channels between them. The channels are buffered to synchronize producers
1417
and consumers to a controllable degree.
1418
1419
### Mailboxes
1420
1421
Flows on different nodes communicate with each other over gRPC streams. To
1422
allow the producer and the consumer to start at different times,
1423
`ScheduleFlows` creates named mailboxes for all the input and output streams.
1424
These message boxes will hold some number of tuples in an internal queue until
1425
a gRPC stream is established for transporting them. From that moment on, gRPC
1426
flow control is used to synchronize the producer and consumer. A gRPC stream is
1427
established by the consumer using the `StreamMailbox` RPC, taking a mailbox id
1428
(the same one that's been already used in the flows passed to `ScheduleFlows`).
1429
1430
A diagram of a simple query using mailboxes for its execution:
1431
1432
1433
## A complex example: Daily Promotion
1434
1435
To give a visual intuition of all the concepts presented, we draw the physical plan of a relatively involved query. The
1436
point of the query is to help with a promotion that goes out daily, targeting
1437
customers that have spent over $1000 in the last year. We'll insert into the
1438
`DailyPromotion` table rows representing each such customer and the sum of her
1439
recent orders.
1440
1441
```SQL
1442
TABLE DailyPromotion (
1443
Email TEXT,
1444
Name TEXT,
1445
OrderCount INT
1446
)
1447
1448
TABLE Customers (
1449
CustomerID INT PRIMARY KEY,
1450
Email TEXT,
1451
Name TEXT
1452
)
1453
1454
TABLE Orders (
1455
CustomerID INT,
1456
Date DATETIME,
1457
Value INT,
1458
1459
PRIMARY KEY (CustomerID, Date),
1460
INDEX date (Date)
1461
)
1462
1463
INSERT INTO DailyPromotion
1464
(SELECT c.Email, c.Name, os.OrderCount FROM
1465
Customers AS c
1466
INNER JOIN
1467
(SELECT CustomerID, COUNT(*) as OrderCount FROM Orders
1468
WHERE Date >= '2015-01-01'
1469
GROUP BY CustomerID HAVING SUM(Value) >= 1000) AS os
1470
ON c.CustomerID = os.CustomerID)
1471
```
1472
1473
A possible physical plan:
1474
