Permalink
Apr 23, 2015
May 21, 2015
Apr 23, 2015
May 29, 2015
Apr 23, 2015
May 22, 2015
Apr 23, 2015
May 22, 2015
Apr 23, 2015
Jun 1, 2015
May 29, 2015
Jun 1, 2015
May 29, 2015
Apr 23, 2015
May 25, 2015
Apr 23, 2015
May 29, 2015
Apr 23, 2015
Apr 23, 2015
May 22, 2015
Apr 23, 2015
May 22, 2015
Apr 23, 2015
May 22, 2015
Apr 23, 2015
Apr 23, 2015
May 21, 2015
Apr 23, 2015
May 22, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
May 22, 2015
May 21, 2015
Apr 23, 2015
May 25, 2015
Apr 23, 2015
May 29, 2015
May 29, 2015
Apr 23, 2015
May 29, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
May 29, 2015
Apr 23, 2015
Apr 23, 2015
May 28, 2015
May 29, 2015
Apr 23, 2015
May 28, 2015
Apr 23, 2015
May 28, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
May 29, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Newer
100644
1290 lines (1070 sloc)
61.4 KB
1
# About
2
This document is an updated version of the original design documents
3
by Spencer Kimball from early 2014.
4
5
# Overview
6
7
Cockroach is a distributed key:value datastore (SQL and structured
8
data layers of cockroach have yet to be defined) which supports **ACID
9
transactional semantics** and **versioned values** as first-class
10
features. The primary design goal is **global consistency and
11
survivability**, hence the name. Cockroach aims to tolerate disk,
12
machine, rack, and even **datacenter failures** with minimal latency
13
disruption and **no manual intervention**. Cockroach nodes are
14
symmetric; a design goal is **homogenous deployment** (one binary) with
15
minimal configuration.
16
17
Cockroach implements a **single, monolithic sorted map** from key to
18
value where both keys and values are byte strings (not unicode).
19
Cockroach **scales linearly** (theoretically up to 4 exabytes (4E) of
20
logical data). The map is composed of one or more ranges and each range
21
is backed by data stored in [RocksDB](http://rocksdb.org/) (a
22
variant of LevelDB), and is replicated to a total of three or more
23
cockroach servers. Ranges are defined by start and end keys. Ranges are
24
merged and split to maintain total byte size within a globally
25
configurable min/max size interval. Range sizes default to target `64M` in
26
order to facilitate quick splits and merges and to distribute load at
27
hotspots within a key range. Range replicas are intended to be located
28
in disparate datacenters for survivability (e.g. `{ US-East, US-West,
29
Japan }`, `{ Ireland, US-East, US-West}`, `{ Ireland, US-East, US-West,
30
Japan, Australia }`).
31
32
Single mutations to ranges are mediated via an instance of a distributed
33
consensus algorithm to ensure consistency. We’ve chosen to use the
34
[Raft consensus
35
algorithm](https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf).
36
All consensus state is stored in RocksDB.
37
38
A single logical mutation may affect multiple key/value pairs. Logical
39
mutations have ACID transactional semantics. If all keys affected by a
40
logical mutation fall within the same range, atomicity and consistency
41
are guaranteed by Raft; this is the **fast commit path**. Otherwise, a
42
**non-locking distributed commit** protocol is employed between affected
43
ranges.
44
45
Cockroach provides [snapshot isolation](http://en.wikipedia.org/wiki/Snapshot_isolation) (SI) and
46
serializable snapshot isolation (SSI) semantics, allowing **externally
47
consistent, lock-free reads and writes**--both from a historical
48
snapshot timestamp and from the current wall clock time. SI provides
49
lock-free reads and writes but still allows write skew. SSI eliminates
50
write skew, but introduces a performance hit in the case of a
51
contentious system. SSI is the default isolation; clients must
52
consciously decide to trade correctness for performance. Cockroach
53
implements [a limited form of linearizability](#linearizability),
54
providing ordering for any observer or chain of observers.
55
56
Similar to
57
[Spanner](http://static.googleusercontent.com/media/research.google.com/en/us/archive/spanner-osdi2012.pdf)
58
directories, Cockroach allows configuration of arbitrary zones of data.
59
This allows replication factor, storage device type, and/or datacenter
60
location to be chosen to optimize performance and/or availability.
61
Unlike Spanner, zones are monolithic and don’t allow movement of fine
62
grained data on the level of entity groups.
63
64
A
65
[Megastore](http://www.cidrdb.org/cidr2011/Papers/CIDR11_Paper32.pdf)-like
66
message queue mechanism is also provided to 1) efficiently sideline
67
updates which can tolerate asynchronous execution and 2) provide an
68
integrated message queuing system for asynchronous communication between
69
distributed system components.
70
71
# Architecture
72
73
Cockroach implements a layered architecture. The highest level of
74
abstraction is the SQL layer (currently unspecified in this document).
75
It depends directly on the [*structured data
76
API*](#structured-data-api), which provides familiar relational concepts
77
such as schemas, tables, columns, and indexes. The structured data API
78
in turn depends on the [distributed key value store](#key-value-api),
79
which handles the details of range addressing to provide the abstraction
80
of a single, monolithic key value store. The distributed KV store
81
communicates with any number of physical cockroach nodes. Each node
82
contains one or more stores, one per physical device.
83
84
85
86
Each store contains potentially many ranges, the lowest-level unit of
87
key-value data. Ranges are replicated using the Raft consensus protocol.
88
The diagram below is a blown up version of stores from four of the five
89
nodes in the previous diagram. Each range is replicated three ways using
90
raft. The color coding shows associated range replicas.
91
92
93
94
Each physical node exports a RoachNode service. Each RoachNode exports
95
one or more key ranges. RoachNodes are symmetric. Each has the same
96
binary and assumes identical roles.
97
98
Nodes and the ranges they provide access to can be arranged with various
99
physical network topologies to make trade offs between reliability and
100
performance. For example, a triplicated (3-way replica) range could have
101
each replica located on different:
102
103
- disks within a server to tolerate disk failures.
104
- servers within a rack to tolerate server failures.
105
- servers on different racks within a datacenter to tolerate rack power/network failures.
106
- servers in different datacenters to tolerate large scale network or power outages.
107
108
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).
109
110
# Cockroach Client
111
112
In order to support diverse client usage, Cockroach clients connect to
113
any node via HTTPS using protocol buffers or JSON. The connected node
114
proxies involved client work including key lookups and write buffering.
115
116
# Keys
117
118
Cockroach keys are arbitrary byte arrays. If textual data is used in
119
keys, utf8 encoding is recommended (this helps for cleaner display of
120
values in debugging tools). User-supplied keys are encoded using an
121
ordered code. System keys are either prefixed with null characters (`\0`
122
or `\0\0`) for system tables, or take the form of
123
`<user-key><system-suffix>` to sort user-key-range specific system
124
keys immediately after the user keys they refer to. Null characters are
125
used in system key prefixes to guarantee that they sort first.
126
127
# Versioned Values
128
129
Cockroach maintains historical versions of values by storing them with
130
associated commit timestamps. Reads and scans can specify a snapshot
131
time to return the most recent writes prior to the snapshot timestamp.
132
Older versions of values are garbage collected by the system during
133
compaction according to a user-specified expiration interval. In order
134
to support long-running scans (e.g. for MapReduce), all versions have a
135
minimum expiration.
136
137
Versioned values are supported via modifications to RocksDB to record
138
commit timestamps and GC expirations per key.
139
144
the low water mark of the cache appropriately. If a new range replica leader
145
is elected, it sets the low water mark for the cache to the current
148
# Lock-Free Distributed Transactions
149
150
Cockroach provides distributed transactions without locks. Cockroach
151
transactions support two isolation levels:
152
153
- snapshot isolation (SI) and
154
- *serializable* snapshot isolation (SSI).
155
156
*SI* is simple to implement, highly performant, and correct for all but a
157
handful of anomalous conditions (e.g. write skew). *SSI* requires just a touch
158
more complexity, is still highly performant (less so with contention), and has
159
no anomalous conditions. Cockroach’s SSI implementation is based on ideas from
160
the literature and some possibly novel insights.
161
162
SSI is the default level, with SI provided for application developers
163
who are certain enough of their need for performance and the absence of
164
write skew conditions to consciously elect to use it. In a lightly
165
contended system, our implementation of SSI is just as performant as SI,
166
requiring no locking or additional writes. With contention, our
167
implementation of SSI still requires no locking, but will end up
168
aborting more transactions. Cockroach’s SI and SSI implementations
169
prevent starvation scenarios even for arbitrarily long transactions.
170
171
See the [Cahill paper](https://drive.google.com/file/d/0B9GCVTp_FHJIcEVyZVdDWEpYYXVVbFVDWElrYUV0NHFhU2Fv/edit?usp=sharing)
172
for one possible implementation of SSI. This is another [great paper](http://cs.yale.edu/homes/thomson/publications/calvin-sigmod12.pdf).
173
For a discussion of SSI implemented by preventing read-write conflicts
174
(in contrast to detecting them, called write-snapshot isolation), see
175
the [Yabandeh paper](https://drive.google.com/file/d/0B9GCVTp_FHJIMjJ2U2t6aGpHLTFUVHFnMTRUbnBwc2pLa1RN/edit?usp=sharing),
176
which is the source of much inspiration for Cockroach’s SSI.
177
178
Each Cockroach transaction is assigned a random priority and a
179
"candidate timestamp" at start. The candidate timestamp is the
180
provisional timestamp at which the transaction will commit, and is
181
chosen as the current clock time of the node coordinating the
182
transaction. This means that a transaction without conflicts will
183
usually commit with a timestamp that, in absolute time, precedes the
184
actual work done by that transaction.
185
186
In the course of coordinating a transaction between one or more
187
distributed nodes, the candidate timestamp may be increased, but will
188
never be decreased. The core difference between the two isolation levels
189
SI and SSI is that the former allows the transaction's candidate
190
timestamp to increase and the latter does not.
191
195
in the [Hybrid Logical Clock
196
paper](http://www.cse.buffalo.edu/tech-reports/2014-04.pdf).
197
HLC time is a combination of both a physical and a
198
logical component to support monotonic increments. We like using
199
the HLC time for cockroach timestamps because it has two important
200
properties:
201
202
For any read/write event x at wall time pt.x, and HLC time hlc.x;
203
204
1. Causality: e happens-before f ⇒ hlc.e < hlc.f.
205
2. hlc.e is close to pt.e, i.e., |hlc.e − pt.e| is bounded.
206
207
happens-before is not happens before in the passing of time sense, but in the
208
passing of information (e.g.: when f reads data written by e). Note, it is
209
not true that hlc.e < hlc.f ⇒ e happens-before f in the time sense
210
(e.g.: when e and f are completely independent). For a more in depth
211
description of HLC please read the paper. Our implementation is
212
[here](https://github.com/cockroachdb/cockroach/blob/master/util/hlc/hlc.go).
213
214
Cockroach picks a Timestamp for a transaction using HLC time. Throughout this
215
document, *timestamp* always refers to the HLC time which is a singleton
216
on each node. The HLC is updated by every read/write event on the node, and
217
the HLC time >= walltime. A read/write timestamp received in a cockroach request
218
from another node is not only used to version the operation, but also updates
219
the HLC on the node. This is useful in guaranteeing that all data read/written
220
on a node is at a timestamp < next HLC time.
221
223
224
1. Start the transaction by writing a new entry to the system
225
transaction table (keys prefixed by *\0tx*) with state “PENDING”. In
226
parallel write an "intent" value for each datum being written as part
227
of the transaction. These are normal MVCC values, with the addition of
228
a special flag (i.e. “intent”) indicating that the value may be
230
the transaction id (unique and chosen at tx start time by client)
231
is stored with intent values. The tx id is used to refer to the
232
transaction table when there are conflicts and to make
233
tie-breaking decisions on ordering between identical timestamps.
235
original candidate timestamp in the absence of read/write conflicts);
236
the client selects the maximum from amongst all write timestamps as the
237
final commit timestamp.
238
240
transaction table (keys prefixed by *\0tx*). The value of the
241
commit entry contains the candidate timestamp (increased as
242
necessary to accommodate any latest read timestamps). Note that
243
the transaction is considered fully committed at this point and
244
control may be returned to the client.
245
246
In the case of an SI transaction, a commit timestamp which was
247
increased to accommodate concurrent readers is perfectly
248
acceptable and the commit may continue. For SSI transactions,
249
however, a gap between candidate and commit timestamps
250
necessitates transaction restart (note: restart is different than
251
abort--see below).
252
253
After the transaction is committed, all written intents are upgraded
254
in parallel by removing the “intent” flag. The transaction is
255
considered fully committed before this step and does not wait for
256
it to return control to the transaction coordinator.
257
258
In the absence of conflicts, this is the end. Nothing else is necessary
259
to ensure the correctness of the system.
260
261
**Conflict Resolution**
262
263
Things get more interesting when a reader or writer encounters an intent
264
record or newly-committed value in a location that it needs to read or
265
write. This is a conflict, usually causing either of the transactions to
266
abort or restart depending on the type of conflict.
267
268
***Transaction restart:***
269
270
This is the usual (and more efficient) type of behaviour and is used
271
except when the transaction was aborted (for instance by another
272
transaction).
273
In effect, that reduces to two cases; the first being the one outlined
274
above: An SSI transaction that finds upon attempting to commit that
275
its commit timestamp has been pushed. The second case involves a transaction
276
actively encountering a conflict, that is, one of its readers or writers
277
encounter data that necessitate conflict resolution
278
(see transaction interactions below).
279
280
When a transaction restarts, it changes its priority and/or moves its
282
begins anew reusing the same tx id. The prior run of the transaction might
283
have written some write intents, which need to be deleted before the
284
transaction commits, so as to not be included as part of the transaction.
285
These stale write intent deletions are done during the reexecution of the
286
transaction, either implicitly, through writing new intents to
287
the same keys as part of the reexecution of the transaction, or explicitly,
288
by cleaning up stale intents that are not part of the reexecution of the
289
transaction. Since most transactions will end up writing to the same keys,
290
the explicit cleanup run just before committing the transaction is usually
291
a NOOP.
292
293
***Transaction abort:***
294
295
This is the case in which a transaction, upon reading its transaction
296
table entry, finds that it has been aborted. In this case, the
297
transaction can not reuse its intents; it returns control to the client
298
before cleaning them up (other readers and writers would clean up
299
dangling intents as they encounter them) but will make an effort to
300
clean up after itself. The next attempt (if applicable) then runs as a
304
305
There are several scenarios in which transactions interact:
306
307
- **Reader encounters write intent or value with newer timestamp far
308
enough in the future**: This is not a conflict. The reader is free
309
to proceed; after all, it will be reading an older version of the
310
value and so does not conflict. Recall that the write intent may
311
be committed with a later timestamp than its candidate; it will
312
never commit with an earlier one. **Side note**: if a SI transaction
313
reader finds an intent with a newer timestamp which the reader’s own
315
316
- **Reader encounters write intent or value with newer timestamp in the
317
near future:** In this case, we have to be careful. The newer
318
intent may, in absolute terms, have happened in our read's past if
319
the clock of the writer is ahead of the node serving the values.
320
In that case, we would need to take this value into account, but
321
we just don't know. Hence the transaction restarts, using instead
322
a future timestamp (but remembering a maximum timestamp used to
323
limit the uncertainty window to the maximum clock skew). In fact,
324
this is optimized further; see the details under "choosing a time
325
stamp" below.
326
327
- **Reader encounters write intent with older timestamp**: the reader
328
must follow the intent’s transaction id to the transaction table.
329
If the transaction has already been committed, then the reader can
330
just read the value. If the write transaction has not yet been
331
committed, then the reader has two options. If the write conflict
332
is from an SI transaction, the reader can *push that transaction's
333
commit timestamp into the future* (and consequently not have to
334
read it). This is simple to do: the reader just updates the
335
transaction’s commit timestamp to indicate that when/if the
336
transaction does commit, it should use a timestamp *at least* as
337
high. However, if the write conflict is from an SSI transaction,
338
the reader must compare priorities. If the reader has the higher priority,
339
it pushes the transaction’s commit timestamp (that
340
transaction will then notice its timestamp has been pushed, and
342
a new priority `max(new random priority, conflicting txn’s
344
346
If the other write intent has been written by a transaction with a lower
347
priority, the writer aborts the conflicting transaction. If the write
348
intent has a higher or equal priority the transaction retries, using as a new
349
priority *max(new random priority, conflicting txn’s priority - 1)*;
350
the retry occurs after a short, randomized backoff interval.
351
352
- **Writer encounters newer committed value**:
353
The committed value could also be an unresolved write intent made by a
354
transaction that has already committed. The transaction restarts. On restart,
355
the same priority is reused, but the candidate timestamp is moved forward
356
to the encountered value's timestamp.
357
360
candidate timestamp is earlier than the low water mark on the cache itself
361
(i.e. its last evicted timestamp) or if the key being written has a read
362
timestamp later than the write’s candidate timestamp, this later timestamp
363
value is returned with the write. A new timestamp forces an SSI transaction
364
to restart.
366
**Transaction management**
367
368
Transactions are managed by the client proxy (or gateway in SQL Azure
369
parlance). Unlike in Spanner, writes are not buffered but are sent
370
directly to all implicated ranges. This allows the transaction to abort
371
quickly if it encounters a write conflict. The client proxy keeps track
372
of all written keys in order to resolve write intents asynchronously upon
373
transaction completion. If a transaction commits successfully, all intents
374
are upgraded to committed. In the event a transaction is aborted, all written
375
intents are deleted. The client proxy doesn’t guarantee it will resolve intents.
376
377
In the event the client proxy restarts before the pending transaction is
379
transaction table until aborted by another transaction. Transactions
380
heartbeat the transaction table every five seconds by default.
381
Transactions encountered by readers or writers with dangling intents
382
which haven’t been heartbeat within the required interval are aborted.
387
388
An exploration of retries with contention and abort times with abandoned
389
transaction is
390
[here](https://docs.google.com/document/d/1kBCu4sdGAnvLqpT-_2vaTbomNmX3_saayWEGYu1j7mQ/edit?usp=sharing).
391
392
**Transaction Table**
393
394
Please see [proto/data.proto](https://github.com/cockroachdb/cockroach/blob/master/proto/data.proto) for the up-to-date structures, the best entry point being `message Transaction`.
395
396
**Pros**
397
398
- No requirement for reliable code execution to prevent stalled 2PC
399
protocol.
400
- Readers never block with SI semantics; with SSI semantics, they may
401
abort.
402
- Lower latency than traditional 2PC commit protocol (w/o contention)
403
because second phase requires only a single write to the
404
transaction table instead of a synchronous round to all
405
transaction participants.
406
- Priorities avoid starvation for arbitrarily long transactions and
407
always pick a winner from between contending transactions (no
408
mutual aborts).
409
- Writes not buffered at client; writes fail fast.
410
- No read-locking overhead required for *serializable* SI (in contrast
411
to other SSI implementations).
412
- Well-chosen (i.e. less random) priorities can flexibly give
413
probabilistic guarantees on latency for arbitrary transactions
414
(for example: make OLTP transactions 10x less likely to abort than
415
low priority transactions, such as asynchronously scheduled jobs).
416
417
**Cons**
418
419
- Reads from non-leader replicas still require a ping to the leader to
421
- Abandoned transactions may block contending writers for up to the
422
heartbeat interval, though average wait is likely to be
423
considerably shorter (see [graph in link](https://docs.google.com/document/d/1kBCu4sdGAnvLqpT-_2vaTbomNmX3_saayWEGYu1j7mQ/edit?usp=sharing)).
424
This is likely considerably more performant than detecting and
425
restarting 2PC in order to release read and write locks.
426
- Behavior different than other SI implementations: no first writer
427
wins, and shorter transactions do not always finish quickly.
428
Element of surprise for OLTP systems may be a problematic factor.
429
- Aborts can decrease throughput in a contended system compared with
430
two phase locking. Aborts and retries increase read and write
431
traffic, increase latency and decrease throughput.
432
433
**Choosing a Timestamp**
434
435
A key challenge of reading data in a distributed system with clock skew
436
is choosing a timestamp guaranteed to be greater than the latest
437
timestamp of any committed transaction (in absolute time). No system can
438
claim consistency and fail to read already-committed data.
439
440
Accomplishing consistency for transactions (or just single operations)
441
accessing a single node is easy. The transaction uses the HLC time as the
442
timestamp which is guaranteed to be at a greater timestamp than all the
443
timestamped data on the node.
444
446
transaction `t` is used. In addition, a maximum timestamp `t+ε` is
447
supplied to provide an upper bound on timestamps for already-committed
448
data (`ε` is the maximum clock skew). As the transaction progresses, any
449
data read which have timestamps greater than `t` but less than `t+ε`
450
cause the transaction to abort and retry with the conflicting timestamp
451
t<sub>c</sub>, where t<sub>c</sub> \> t. The maximum timestamp `t+ε` remains
452
the same.
453
457
of the uncertain read t<sub>node<sub>. The larger of those two timestamps
458
(more likely the latter): max(t<sub>c<sub>, t<sub>node<sub>) is used
459
to bump up the read timestamp. Additionally, the conflicting node is
460
marked as “certain”. This means that for future reads to that node
461
within the transaction, we can set `MaxTimestamp = Read Timestamp`.
462
Correctness follows from the fact that we know that at the time of the read,
463
there exists no version of any key on that node with a higher timestamp than
464
t<sub>node<sub>. Upon a restart caused by the node, if the transaction were to
465
encounter a key with a higher timestamp it would imply that the value
466
is written in the future in absolute time, and the transaction can move
467
forward reading an older version of the data (at the transactions timestamp).
468
This limits the time uncertainty restarts attributed to a node to <= 1. The
469
tradeoff is that we might pick a timestamp larger than the optimal one
470
(> highest conflicting timestamp), resulting in the possibility of a few
471
more conflicts.
472
473
We expect retries will be rare, but this assumption may need to be
474
revisited if retries become problematic. Note that this problem does not
475
apply to historical reads. An alternate approach which does not require
476
retries makes a round to all node participants in advance and
477
chooses the highest reported node wall time as the timestamp. However,
478
knowing which nodes will be accessed in advance is difficult and
479
potentially limiting. Cockroach could also potentially use a global
480
clock (Google did this with [Percolator](https://www.usenix.org/legacy/event/osdi10/tech/full_papers/Peng.pdf)), which would be
481
feasible for smaller, geographically-proximate clusters.
482
483
# Linearizability
484
485
First a word about [***Spanner***](http://research.google.com/archive/spanner.html).
486
By combining judicious use of wait intervals with accurate time signals,
487
Spanner provides a global ordering between any two non-overlapping transactions
488
(in absolute time) with \~14ms latencies. Put another way:
489
Spanner guarantees that if a transaction T<sub>1</sub> commits (in absolute time)
490
before another transaction T<sub>2</sub> starts, then T<sub>1</sub>'s assigned commit
491
timestamp is smaller than T<sub>2</sub>'s. Using atomic clocks and GPS receivers,
492
Spanner reduces their clock skew uncertainty to \< 10ms (`ε`). To make
493
good on the promised guarantee, transactions must take at least double
494
the clock skew uncertainty interval to commit (`2ε`). See [*this
495
article*](http://www.cs.cornell.edu/~ie53/publications/DC-col51-Sep13.pdf)
496
for a helpful overview of Spanner’s concurrency control.
497
498
Cockroach could make the same guarantees without specialized hardware,
499
at the expense of longer wait times. If servers in the cluster were
500
configured to work only with NTP, transaction wait times would likely to
501
be in excess of 150ms. For wide-area zones, this would be somewhat
502
mitigated by overlap from cross datacenter link latencies. If clocks
503
were made more accurate, the minimal limit for commit latencies would
504
improve.
505
506
However, let’s take a step back and evaluate whether Spanner’s external
507
consistency guarantee is worth the automatic commit wait. First, if the
508
commit wait is omitted completely, the system still yields a consistent
509
view of the map at an arbitrary timestamp. However with clock skew, it
510
would become possible for commit timestamps on non-overlapping but
511
causally related transactions to suffer temporal reverse. In other
512
words, the following scenario is possible for a client without global
513
ordering:
514
515
- Start transaction T<sub>1</sub> to modify value `x` with commit time *s<sub>1</sub>*
516
517
- On commit of T<sub>1</sub>, start T<sub>2</sub> to modify value `y` with commit time
518
\> s<sub>2</sub>
519
520
- Read `x` and `y` and discover that s<sub>1</sub> \> s<sub>2</sub> (**!**)
521
522
The external consistency which Spanner guarantees is referred to as
523
**linearizability**. It goes beyond serializability by preserving
524
information about the causality inherent in how external processes
525
interacted with the database. The strength of Spanner’s guarantee can be
526
formulated as follows: any two processes, with clock skew within
527
expected bounds, may independently record their wall times for the
528
completion of transaction T<sub>1</sub> (T<sub>1</sub><sup>end</sup>) and start of transaction
529
T<sub>2</sub> (T<sub>2</sub><sup>start</sup>) respectively, and if later
530
compared such that T<sub>1</sub><sup>end</sup> \< T<sub>2</sub><sup>start</sup>,
531
then commit timestamps s<sub>1</sub> \< s<sub>2</sub>.
532
This guarantee is broad enough to completely cover all cases of explicit
533
causality, in addition to covering any and all imaginable scenarios of implicit
534
causality.
535
536
Our contention is that causality is chiefly important from the
537
perspective of a single client or a chain of successive clients (*if a
538
tree falls in the forest and nobody hears…*). As such, Cockroach
539
provides two mechanisms to provide linearizability for the vast majority
540
of use cases without a mandatory transaction commit wait or an elaborate
541
system to minimize clock skew.
542
543
1. Clients provide the highest transaction commit timestamp with
544
> successive transactions. This allows node clocks from previous
545
> transactions to effectively participate in the formulation of the
546
> commit timestamp for the current transaction. This guarantees
547
> linearizability for transactions committed by this client.
548
>
549
> Newly launched clients wait at least 2 \* ε from process start
550
> time before beginning their first transaction. This preserves the
551
> same property even on client restart, and the wait will be
552
> mitigated by process initialization.
553
>
554
> All causally-related events within Cockroach maintain
555
> linearizability. Message queues, for example, guarantee that the
556
> receipt timestamp is greater than send timestamp, and that
557
> delivered messages may not be reaped until after the commit wait.
558
559
2. Committed transactions respond with a commit wait parameter which
560
> represents the remaining time in the nominal commit wait. This
561
> will typically be less than the full commit wait as the consensus
562
> write at the coordinator accounts for a portion of it.
563
>
564
> Clients taking any action outside of another Cockroach transaction
565
> (e.g. writing to another distributed system component) can either
566
> choose to wait the remaining interval before proceeding, or
567
> alternatively, pass the wait and/or commit timestamp to the
568
> execution of the outside action for its consideration. This pushes
569
> the burden of linearizability to clients, but is a useful tool in
570
> mitigating commit latencies if the clock skew is potentially
571
> large. This functionality can be used for ordering in the face of
572
> backchannel dependencies as mentioned in the
573
> [AugmentedTime](http://www.cse.buffalo.edu/~demirbas/publications/augmentedTime.pdf)
574
> paper.
575
576
Using these mechanisms in place of commit wait, Cockroach’s guarantee can be
577
formulated as follows: any process which signals the start of transaction
578
T<sub>2</sub> (T<sub>2</sub><sup>start</sup>) after the completion of
579
transaction T<sub>1</sub> (T<sub>1</sub><sup>end</sup>), will have commit
580
timestamps such thats<sub>1</sub> \< s<sub>2</sub>.
581
582
# Logical Map Content
583
584
Logically, the map contains a series of reserved system key / value
585
pairs covering accounting, range metadata, node accounting and
586
permissions before the actual key / value pairs for non-system data
587
(e.g. the actual meat of the map).
588
589
- `\0\0meta1` Range metadata for location of `\0\0meta2`.
590
- `\0\0meta1<key1>` Range metadata for location of `\0\0meta2<key1>`.
591
- ...
592
- `\0\0meta1<keyN>`: Range metadata for location of `\0\0meta2<keyN>`.
593
- `\0\0meta2`: Range metadata for location of first non-range metadata key.
594
- `\0\0meta2<key1>`: Range metadata for location of `<key1>`.
595
- ...
596
- `\0\0meta2<keyN>`: Range metadata for location of `<keyN>`.
597
- `\0acct<key0>`: Accounting for key prefix key0.
598
- ...
599
- `\0acct<keyN>`: Accounting for key prefix keyN.
600
- `\0node<node-address0>`: Accounting data for node 0.
601
- ...
602
- `\0node<node-addressN>`: Accounting data for node N.
603
- `\0perm<key0><user0>`: Permissions for user0 for key prefix key0.
604
- ...
605
- `\0perm<keyN><userN>`: Permissions for userN for key prefix keyN.
606
- `\0tree_root`: Range key for root of range-spanning tree.
607
- `\0tx<tx-id0>`: Transaction record for transaction 0.
608
- ...
609
- `\0tx<tx-idN>`: Transaction record for transaction N.
610
- `\0zone<key0>`: Zone information for key prefix key0.
611
- ...
612
- `\0zone<keyN>`: Zone information for key prefix keyN.
613
- `<>acctd<metric0>`: Accounting data for Metric 0 for empty key prefix.
614
- ...
615
- `<>acctd<metricN>`: Accounting data for Metric N for empty key prefix.
616
- `<key0>`: `<value0>` The first user data key.**
617
- ...
618
- `<keyN>`: `<valueN>` The last user data key.**
619
620
There are some additional system entries sprinkled amongst the
621
non-system keys. See the Key-Prefix Accounting section in this document
622
for further details.
623
624
# Node Storage
625
626
Nodes maintain a separate instance of RocksDB for each disk. Each
627
RocksDB instance hosts any number of ranges. RPCs arriving at a
628
RoachNode are multiplexed based on the disk name to the appropriate
629
RocksDB instance. A single instance per disk is used to avoid
630
contention. If every range maintained its own RocksDB, global management
631
of available cache memory would be impossible and writers for each range
632
would compete for non-contiguous writes to multiple RocksDB logs.
633
634
In addition to the key/value pairs of the range itself, various range
635
metadata is maintained.
636
637
- range-spanning tree node links
638
639
- participating replicas
640
641
- consensus metadata
642
643
- split/merge activity
644
645
A really good reference on tuning Linux installations with RocksDB is
646
[here](http://docs.basho.com/riak/latest/ops/advanced/backends/leveldb/).
647
648
# Range Metadata
649
650
The default approximate size of a range is 64M (2\^26 B). In order to
651
support 1P (2\^50 B) of logical data, metadata is needed for roughly
652
2\^(50 - 26) = 2\^24 ranges. A reasonable upper bound on range metadata
654
locations and 220 bytes for the range key itself*). 2\^24 ranges \* 2\^8
655
B would require roughly 4G (2\^32 B) to store--too much to duplicate
656
between machines. Our conclusion is that range metadata must be
657
distributed for large installations.
658
659
To distribute the range metadata and keep key lookups relatively fast,
660
we use two levels of indirection. All of the range metadata sorts first
661
in our key-value map. We accomplish this by prefixing range metadata
662
with two null characters (*\0\0*). The *meta1* or *meta2* suffixes are
663
additionally appended to distinguish between the first level and second
665
we first locate the range information for the lower bound of
666
`\0\0meta1<key1>`, and then use that range to locate the lower bound
667
of `\0\0meta2<key1>`. The range specified there will indicate the
668
range location of `<key1>` (refer to examples below). Using two levels
669
of indirection, **our map can address approximately 2\^62 B of data, or
670
roughly 4E** (*each metadata range addresses 2\^(26-8) = 2\^18 ranges;
671
with two levels of indirection, we can address 2\^(18 + 18) = 2\^36
672
ranges; each range addresses 2\^26 B; total is 2\^(36+26) B = 2\^62 B =
673
4E*).
674
675
Note: we append the end key of each range to meta[12] records because
676
the RocksDB iterator only supports a Seek() interface which acts as a
677
Ceil(). Using the start key of the range would cause Seek() to find the
678
key *after* the meta indexing record we’re looking for, which would
679
result in having to back the iterator up, an option which is both less
680
efficient and not available in all cases.
681
682
The following example shows the directory structure for a map with
683
three ranges worth of data. Ellipses indicate additional key/value pairs to
684
fill an entire range of data. Except for the fact that splitting ranges
685
requires updates to the range metadata with knowledge of the metadata layout,
686
the range metadata itself requires no special treatment or bootstrapping.
687
688
**Range 0** (located on servers `dcrama1:8000`, `dcrama2:8000`,
689
`dcrama3:8000`)
690
691
- `\0\0meta1\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
692
- `\0\0meta2<lastkey0>`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
693
- `\0\0meta2<lastkey1>`: `dcrama4:8000`, `dcrama5:8000`, `dcrama6:8000`
694
- `\0\0meta2\xff`: `dcrama7:8000`, `dcrama8:8000`, `dcrama9:8000`
695
- ...
696
- `<lastkey0>`: `<lastvalue0>`
697
698
**Range 1** (located on servers `dcrama4:8000`, `dcrama5:8000`,
699
`dcrama6:8000`)
700
701
- ...
702
- `<lastkey1>`: `<lastvalue1>`
703
704
**Range 2** (located on servers `dcrama7:8000`, `dcrama8:8000`,
705
`dcrama9:8000`)
706
707
- ...
708
- `<lastkey2>`: `<lastvalue2>`
709
710
Consider a simpler example of a map containing less than a single
711
range of data. In this case, all range metadata and all data are
712
located in the same range:
713
714
**Range 0** (located on servers `dcrama1:8000`, `dcrama2:8000`,
715
`dcrama3:8000`)*
716
717
- `\0\0meta1\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
718
- `\0\0meta2\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
719
- `<key0>`: `<value0>`
720
- `...`
721
722
Finally, a map large enough to need both levels of indirection would
723
look like (note that instead of showing range replicas, this
724
example is simplified to just show range indexes):
725
726
**Range 0**
727
728
- `\0\0meta1<lastkeyN-1>`: Range 0
729
- `\0\0meta1\xff`: Range 1
730
- `\0\0meta2<lastkey1>`: Range 1
731
- `\0\0meta2<lastkey2>`: Range 2
732
- `\0\0meta2<lastkey3>`: Range 3
733
- ...
734
- `\0\0meta2<lastkeyN-1>`: Range 262143
735
736
**Range 1**
737
738
- `\0\0meta2<lastkeyN>`: Range 262144
739
- `\0\0meta2<lastkeyN+1>`: Range 262145
740
- ...
741
- `\0\0meta2\xff`: Range 500,000
742
- ...
743
- `<lastkey1>`: `<lastvalue1>`
744
745
**Range 2**
746
747
- ...
748
- `<lastkey2>`: `<lastvalue2>`
749
750
**Range 3**
751
752
- ...
753
- `<lastkey3>`: `<lastvalue3>`
754
755
**Range 262144**
756
757
- ...
758
- `<lastkeyN>`: `<lastvalueN>`
759
760
**Range 262145**
761
762
- ...
763
- `<lastkeyN+1>`: `<lastvalueN+1>`
764
765
Note that the choice of range `262144` is just an approximation. The
766
actual number of ranges addressable via a single metadata range is
767
dependent on the size of the keys. If efforts are made to keep key sizes
768
small, the total number of addressable ranges would increase and vice
769
versa.
770
771
From the examples above it’s clear that key location lookups require at
772
most three reads to get the value for `<key>`:
773
774
1. lower bound of `\0\0meta1<key>`
775
2. lower bound of `\0\0meta2<key>`,
776
3. `<key>`.
777
778
For small maps, the entire lookup is satisfied in a single RPC to Range 0. Maps
779
containing less than 16T of data would require two lookups. Clients cache both
780
levels of range metadata, and we expect that data locality for individual
781
clients will be high. Clients may end up with stale cache entries. If on a
782
lookup, the range consulted does not match the client’s expectations, the
783
client evicts the stale entries and possibly does a new lookup.
784
785
# Range-Spanning Binary Tree
786
787
A crucial enhancement to the organization of range metadata is to
788
augment the bi-level range metadata lookup with a minimum spanning tree,
789
implemented as a left-leaning red-black tree over all ranges in the map.
790
This tree structure allows the system to start at any key prefix and
791
efficiently traverse an arbitrary key range with minimal RPC traffic,
792
minimal fan-in and fan-out, and with bounded time complexity equal to
793
`2*log N` steps, where `N` is the total number of ranges in the system.
794
795
Unlike the range metadata rows prefixed with `\0\0meta[1|2]`, the
796
metadata for the range-spanning tree (e.g. parent range and left / right
797
child ranges) is stored directly at the ranges as non-map metadata. The
798
metadata for each node of the tree (e.g. links to parent range, left
799
child range, and right child range) is stored with the range metadata.
800
In effect, the tree metadata is stored implicitly. In order to traverse
801
the tree, for example, you’d need to query each range in turn for its
802
metadata.
803
804
Any time a range is split or merged, both the bi-level range lookup
805
metadata and the per-range binary tree metadata are updated as part of
806
the same distributed transaction. The total number of nodes involved in
807
the update is bounded by 2 + log N (i.e. 2 updates for meta1 and
808
meta2, and up to log N updates to balance the range-spanning tree).
809
The range corresponding to the root node of the tree is stored in
811
812
As an example, consider the following set of nine ranges and their
813
associated range-spanning tree:
814
815
R0: `aa - cc`, R1: `*cc - lll`, R2: `*lll - llr`, R3: `*llr - nn`, R4: `*nn - rr`, R5: `*rr - ssss`, R6: `*ssss - sst`, R7: `*sst - vvv`, R8: `*vvv - zzzz`.
816
817
818
819
The range-spanning tree has many beneficial uses in Cockroach. It makes
820
the problem of efficiently aggregating accounting information of
821
potentially vast ranges of data tractable. Imagine a subrange of data
822
over which accounting is being kept. For example, the *photos* table in
823
a public photo sharing site. To efficiently keep track of data about the
824
table (e.g. total size, number of rows, etc.), messages can be passed
825
first up the tree and then down to the left until updates arrive at the
826
key prefix under which accounting is aggregated. This makes worst case
827
number of hops for an update to propagate into the accounting totals
828
2 \* log N. A 64T database will require 1M ranges, meaning 40 hops
829
worst case. In our experience, accounting tasks over vast ranges of data
830
are most often map/reduce jobs scheduled with coarse-grained
831
periodicity. By contrast, we expect Cockroach to maintain statistics
832
with sub 10s accuracy and with minimal cycles and minimal IOPs.
833
834
Another use for the range-spanning tree is to push accounting, zones and
835
permissions configurations to all ranges. In the case of zones and
836
permissions, this is an efficient way to pass updated configuration
837
information with exponential fan-out. When adding accounting
838
configurations (i.e. specifying a new key prefix to track), the
839
implicated ranges are transactionally scanned and zero-state accounting
840
information is computed as well. Deleting accounting configurations is
841
similar, except accounting records are deleted.
842
843
Last but *not* least, the range-spanning tree provides a convenient
844
mechanism for planning and executing parallel queries. These provide the
845
basis for
846
[Dremel](http://static.googleusercontent.com/media/research.google.com/en/us/pubs/archive/36632.pdf)-like
847
query execution trees and it’s easy to imagine supporting a subset of
848
SQL or even javascript-based user functions for complex data analysis
849
tasks.
850
851
# Raft - Consistency of Range Replicas
852
853
Each range is configured to consist of three or more replicas. The
854
replicas in a range maintain their own instance of a distributed
855
consensus algorithm. We use the [*Raft consensus
856
algorithm*](https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf)
857
as it is simpler to reason about and includes a reference implementation
858
covering important details. Every write to replicas is logged twice.
859
Once to RocksDB’s internal log and once to levedb itself as part of the
860
Raft consensus log.
861
[ePaxos](https://www.cs.cmu.edu/~dga/papers/epaxos-sosp2013.pdf) has
862
promising performance characteristics for WAN-distributed replicas, but
863
it does not guarantee a consistent ordering between replicas.
864
865
Raft elects a relatively long-lived leader which must be involved to
866
propose writes. It heartbeats followers periodically to keep their logs
867
replicated. In the absence of heartbeats, followers become candidates
868
after randomized election timeouts and proceed to hold new leader
869
elections. Cockroach weights random timeouts such that the replicas with
870
shorter round trip times to peers are more likely to hold elections
871
first. Although only the leader can propose a new write, and as such
872
must be involved in any write to the consensus log, any replica can
873
service reads if the read is for a timestamp which the replica knows is
874
safe based on the last committed consensus write and the state of any
875
pending transactions.
876
877
Only the leader can propose a new write, but Cockroach accepts writes at
878
any replica. The replica merely forwards the write to the leader.
879
Instead of resending the write, the leader has only to acknowledge the
880
write to the forwarding replica using a log sequence number, as though
881
it were proposing it in the first place. The other replicas receive the
882
full write as though the leader were the originator.
883
884
Having a stable leader provides the choice of replica to handle
885
range-specific maintenance and processing tasks, such as delivering
886
pending message queues, handling splits and merges, rebalancing, etc.
887
888
# Splitting / Merging Ranges
889
890
RoachNodes split or merge ranges based on whether they exceed maximum or
891
minimum thresholds for capacity or load. Ranges exceeding maximums for
892
either capacity or load are split; ranges below minimums for *both*
893
capacity and load are merged.
894
895
Ranges maintain the same accounting statistics as accounting key
896
prefixes. These boil down to a time series of data points with minute
897
granularity. Everything from number of bytes to read/write queue sizes.
898
Arbitrary distillations of the accounting stats can be determined as the
899
basis for splitting / merging. Two sensical metrics for use with
900
split/merge are range size in bytes and IOps. A good metric for
901
rebalancing a replica from one node to another would be total read/write
902
queue wait times. These metrics are gossipped, with each range / node
903
passing along relevant metrics if they’re in the bottom or top of the
904
range it’s aware of.
905
906
A range finding itself exceeding either capacity or load threshold
907
splits. To this end, the range leader computes an appropriate split key
908
candidate and issues the split through Raft. In contrast to splitting,
909
merging requires a range to be below the minimum threshold for both
910
capacity *and* load. A range being merged chooses the smaller of the
911
ranges immediately preceding and succeeding it.
912
913
Splitting, merging, rebalancing and recovering all follow the same basic
914
algorithm for moving data between roach nodes. New target replicas are
915
created and added to the replica set of source range. Then each new
916
replica is brought up to date by either replaying the log in full or
917
copying a snapshot of the source replica data and then replaying the log
918
from the timestamp of the snapshot to catch up fully. Once the new
919
replicas are fully up to date, the range metadata is updated and old,
920
source replica(s) deleted if applicable.
921
922
**Coordinator** (leader replica)
923
927
only after being completed locally, are moved to new target replicas.
928
else if merging
929
Choose new replicas on same servers as target range replicas;
930
add to replica set.
931
else if rebalancing || recovering
932
Choose new replica(s) on least loaded servers; add to replica set.
933
```
934
935
**New Replica**
936
937
*Bring replica up to date:*
938
939
```
940
if all info can be read from replicated log
941
copy replicated log
942
else
943
snapshot source replica
944
send successive ReadRange requests to source replica
945
referencing snapshot
946
947
if merging
948
combine ranges on all replicas
949
else if rebalancing || recovering
950
remove old range replica(s)
951
```
952
953
RoachNodes split ranges when the total data in a range exceeds a
954
configurable maximum threshold. Similarly, ranges are merged when the
955
total data falls below a configurable minimum threshold.
956
957
**TBD: flesh this out**.
958
959
Ranges are rebalanced if a node determines its load or capacity is one
960
of the worst in the cluster based on gossipped load stats. A node with
961
spare capacity is chosen in the same datacenter and a special-case split
962
is done which simply duplicates the data 1:1 and resets the range
963
configuration metadata.
964
965
# Message Queues
966
967
Each range maintains an array of incoming message queues, referred to
968
here as **inboxes**. Additionally, each range maintains and *processes*
969
an array of outgoing message queues, referred to here as **outboxes**.
970
Both inboxes and outboxes are assigned to keys; messages can be sent or
971
received on behalf of any key. Inboxes and outboxes can contain any
972
number of pending messages.
973
974
Messages can be *deliverable*, or *executable.*
975
976
Deliverable messages are defined by Value objects - simple byte arrays -
977
that are delivered to a key’s inbox, awaiting collection by a client
978
invoking the ReapQueue operation. These are typically used by client
979
applications wishing to be notified of changes to an entry for further
980
processing, such as expensive offline operations like sending emails,
981
SMSs, etc.
982
983
Executable messages are *outgoing-only*, and are instances of
984
PutRequest,IncrementRequest, DeleteRequest, DeleteRangeRequest
986
executed when encountered. These are primarily useful when updates that
987
are nominally part of a transaction can tolerate asynchronous execution
988
(e.g. eventual consistency), and are otherwise too busy or numerous to
989
make including them in the original [distributed] transaction efficient.
990
Examples may include updates to the accounting for successive key
991
prefixes (potentially busy) or updates to a full-text index (potentially
992
numerous).
993
994
These two types of messages are enqueued in different outboxes too - see
995
key formats below.
996
997
At commit time, the range processing the transaction places messages
998
into a shared outbox located at the start of the range metadata. This is
999
effectively free as it’s part of the same consensus write for the range
1000
as the COMMIT record. Outgoing messages are processed asynchronously by
1001
the range. To make processing easy, all outboxes are co-located at the
1002
start of the range. To make lookup easy, all inboxes are located
1003
immediately after the recipient key. The leader replica of a range is
1004
responsible for processing message queues.
1005
1006
A dispatcher polls a given range’s *deliverable message outbox*
1007
periodically (configurable), and delivers those messages to the target
1008
key’s inbox. The dispatcher is also woken up whenever a new message is
1009
added to the outbox. A separate executor also polls the range’s
1010
*executable message outbox* periodically as well (again, configurable),
1012
new message is added to the outbox.
1013
1014
Formats follow in the table below. Notice that inbox messages for a
1015
given key sort by the `<outbox-timestamp>`. This doesn’t provide a
1016
precise ordering, but it does allow clients to scan messages in an
1017
approximate ordering of when they were originally lodged with senders.
1018
NTP offers walltime deltas to within 100s of milliseconds. The
1019
`<sender-range-key>` suffix provides uniqueness.
1020
1021
**Outbox**
1022
`<sender-range-key>deliverable-outbox:<recipient-key><outbox-timestamp>`
1023
`<sender-range-key>executable-outbox:<recipient-key><outbox-timestamp>`
1024
1025
**Inbox**
1026
`<recipient-key>inbox:<outbox-timestamp><sender-range-key>`
1027
1028
Messages are processed and then deleted as part of a single distributed
1029
transaction. The message will be executed or delivered exactly once,
1030
regardless of failures at either sender or receiver.
1031
1032
Delivered messages may be read by clients via the ReapQueue operation.
1033
This operation may only be used as part of a transaction. Clients should
1034
commit only after having processed the message. If the transaction is
1035
committed, scanned messages are automatically deleted. The operation
1036
name was chosen to reflect its mutating side effect. Deletion of read
1037
messages is mandatory because senders deliver messages asynchronously
1038
and a delay could cause insertion of messages at arbitrary points in the
1039
inbox queue. If clients require persistence, they should re-save read
1040
messages manually; the ReapQueue operation can be incorporated into
1041
normal transactional updates.
1042
1043
# Node Allocation (via Gossip)
1044
1045
New nodes must be allocated when a range is split. Instead of requiring
1046
every RoachNode to know about the status of all or even a large number
1047
of peer nodes --or-- alternatively requiring a specialized curator or
1048
master with sufficiently global knowledge, we use a gossip protocol to
1049
efficiently communicate only interesting information between all of the
1050
nodes in the cluster. What’s interesting information? One example would
1051
be whether a particular node has a lot of spare capacity. Each node,
1052
when gossiping, compares each topic of gossip to its own state. If its
1053
own state is somehow “more interesting” than the least interesting item
1054
in the topic it’s seen recently, it includes its own state as part of
1055
the next gossip session with a peer node. In this way, a node with
1056
capacity sufficiently in excess of the mean quickly becomes discovered
1057
by the entire cluster. To avoid piling onto outliers, nodes from the
1058
high capacity set are selected at random for allocation.
1059
1060
The gossip protocol itself contains two primary components:
1061
1062
- **Peer Selection**: each node maintains up to N peers with which it
1063
regularly communicates. It selects peers with an eye towards
1064
maximizing fanout. A peer node which itself communicates with an
1065
array of otherwise unknown nodes will be selected over one which
1066
communicates with a set containing significant overlap. Each time
1067
gossip is initiated, each nodes’ set of peers is exchanged. Each
1068
node is then free to incorporate the other’s peers as it sees fit.
1069
To avoid any node suffering from excess incoming requests, a node
1070
may refuse to answer a gossip exchange. Each node is biased
1071
towards answering requests from nodes without significant overlap
1072
and refusing requests otherwise.
1073
1074
Peers are efficiently selected using a heuristic as described in
1075
[Agarwal & Trachtenberg (2006)](https://drive.google.com/file/d/0B9GCVTp_FHJISmFRTThkOEZSM1U/edit?usp=sharing).
1076
1077
**TBD**: how to avoid partitions? Need to work out a simulation of
1078
the protocol to tune the behavior and see empirically how well it
1079
works.
1080
1081
- **Gossip Selection**: what to communicate. Gossip is divided into
1082
topics. Load characteristics (capacity per disk, cpu load, and
1083
state [e.g. draining, ok, failure]) are used to drive node
1084
allocation. Range statistics (range read/write load, missing
1085
replicas, unavailable ranges) and network topology (inter-rack
1086
bandwidth/latency, inter-datacenter bandwidth/latency, subnet
1087
outages) are used for determining when to split ranges, when to
1088
recover replicas vs. wait for network connectivity, and for
1089
debugging / sysops. In all cases, a set of minimums and a set of
1090
maximums is propagated; each node applies its own view of the
1091
world to augment the values. Each minimum and maximum value is
1092
tagged with the reporting node and other accompanying contextual
1093
information. Each topic of gossip has its own protobuf to hold the
1094
structured data. The number of items of gossip in each topic is
1095
limited by a configurable bound.
1096
1097
For efficiency, nodes assign each new item of gossip a sequence
1098
number and keep track of the highest sequence number each peer
1099
node has seen. Each round of gossip communicates only the delta
1100
containing new items.
1101
1102
# Node Accounting
1103
1104
The gossip protocol discussed in the previous section is useful to
1105
quickly communicate fragments of important information in a
1106
decentralized manner. However, complete accounting for each node is also
1107
stored to a central location, available to any dashboard process. This
1108
is done using the map itself. Each node periodically writes its state to
1109
the map with keys prefixed by `\0node`, similar to the first level of
1110
range metadata, but with an ‘`node`’ suffix. Each value is a protobuf
1111
containing the full complement of node statistics--everything
1112
communicated normally via the gossip protocol plus other useful, but
1113
non-critical data.
1114
1115
The range containing the first key in the node accounting table is
1116
responsible for gossiping the total count of nodes. This total count is
1117
used by the gossip network to most efficiently organize itself. In
1118
particular, the maximum number of hops for gossipped information to take
1119
before reaching a node is given by `ceil(log(node count) / log(max
1120
fanout)) + 1`.
1121
1122
# Key-prefix Accounting, Zones & Permissions
1123
1124
Arbitrarily fine-grained accounting and permissions are specified via
1125
key prefixes. Key prefixes can overlap, as is necessary for capturing
1126
hierarchical relationships. For illustrative purposes, let’s say keys
1127
specifying rows in a set of databases have the following format:
1128
1129
`<db>:<table>:<primary-key>[:<secondary-key>]`
1130
1131
In this case, we might collect accounting or specify permissions with
1132
key prefixes:
1133
1134
`db1`, `db1:user`, `db1:order`,
1135
1136
Accounting is kept for the entire map by default.
1137
1138
## Accounting
1139
to keep accounting for a range defined by a key prefix, an entry is created in
1140
the accounting system table. The format of accounting table keys is:
1141
1142
`\0acct<key-prefix>`
1143
1144
In practice, we assume each RoachNode capable of caching the
1145
entire accounting table as it is likely to be relatively small.
1146
1147
Accounting is kept for key prefix ranges with eventual consistency
1148
for efficiency. Updates to accounting values propagate through the
1149
system using the message queue facility if the accounting keys do
1150
not reside on the same range as ongoing activity (true for all but
1151
the smallest systems). There are two types of values which
1152
comprise accounting: counts and occurrences, for lack of better
1153
terms. Counts describe system state, such as the total number of
1154
bytes, rows, etc. Occurrences include transient performance and
1155
load metrics. Both types of accounting are captured as time series
1156
with minute granularity. The length of time accounting metrics are
1157
kept is configurable. Below are examples of each type of
1158
accounting value.
1159
1160
**System State Counters/Performance**
1161
1162
- Count of items (e.g. rows)
1163
- Total bytes
1164
- Total key bytes
1165
- Total value length
1166
- Queued message count
1167
- Queued message total bytes
1168
- Count of values \< 16B
1169
- Count of values \< 64B
1170
- Count of values \< 256B
1171
- Count of values \< 1K
1172
- Count of values \< 4K
1173
- Count of values \< 16K
1174
- Count of values \< 64K
1175
- Count of values \< 256K
1176
- Count of values \< 1M
1177
- Count of values \> 1M
1178
- Total bytes of accounting
1179
1180
1181
**Load Occurences**
1182
1183
Get op count
1184
Get total MB
1185
Put op count
1186
Put total MB
1187
Delete op count
1188
Delete total MB
1189
Delete range op count
1190
Delete range total MB
1191
Scan op count
1192
Scan op MB
1193
Split count
1194
Merge count
1195
1196
Because accounting information is kept as time series and over many
1197
possible metrics of interest, the data can become numerous. Accounting
1198
data are stored in the map near the key prefix described, in order to
1199
distribute load (for both aggregation and storage).
1200
1201
Accounting keys for system state have the form:
1202
`<key-prefix>|acctd<metric-name>*`. Notice the leading ‘pipe’
1203
character. It’s meant to sort the root level account AFTER any other
1204
system tables. They must increment the same underlying values as they
1205
are permanent counts, and not transient activity. Logic at the
1206
RoachNode takes care of snapshotting the value into an appropriately
1207
suffixed (e.g. with timestamp hour) multi-value time series entry.
1208
1209
Keys for perf/load metrics:
1210
`<key-prefix>acctd<metric-name><hourly-timestamp>`.
1211
1212
`<hourly-timestamp>`-suffixed accounting entries are multi-valued,
1213
containing a varint64 entry for each minute with activity during the
1214
specified hour.
1215
1216
To efficiently keep accounting over large key ranges, the task of
1217
aggregation must be distributed. If activity occurs within the same
1218
range as the key prefix for accounting, the updates are made as part
1219
of the consensus write. If the ranges differ, then a message is sent
1220
to the parent range to increment the accounting. If upon receiving the
1221
message, the parent range also does not include the key prefix, it in
1222
turn forwards it to its parent or left child in the balanced binary
1223
tree which is maintained to describe the range hierarchy. This limits
1224
the number of messages before an update is visible at the root to `2*log N`,
1225
where `N` is the number of ranges in the key prefix.
1226
1227
## Zones
1228
zones are stored in the map with keys prefixed by
1229
`\0zone` followed by the key prefix to which the zone
1230
configuration applies. Zone values specify a protobuf containing
1231
the datacenters from which replicas for ranges which fall under
1232
the zone must be chosen.
1233
1234
Please see [proto/config.proto](https://github.com/cockroachdb/cockroach/blob/master/proto/config.proto) for up-to-date data structures used, the best entry point being `message ZoneConfig`.
1235
1236
If zones are modified in situ, each RoachNode verifies the
1237
existing zones for its ranges against the zone configuration. If
1238
it discovers differences, it reconfigures ranges in the same way
1239
that it rebalances away from busy nodes, via special-case 1:1
1240
split to a duplicate range comprising the new configuration.
1241
1242
### Permissions
1243
permissions are stored in the map with keys prefixed by *\0perm* followed by
1244
the key prefix and user to which the specified permissions apply. The format of
1245
permissions keys is:
1246
1247
`\0perm<key-prefix><user>`
1248
1249
Permission values are a protobuf containing the permission details;
1250
please see [proto/config.proto](https://github.com/cockroachdb/cockroach/blob/master/proto/config.proto) for up-to-date data structures used, the best entry point being `message PermConfig`.
1251
1252
A default system root permission is assumed for the entire map
1253
with an empty key prefix and read/write as true.
1254
1255
When determining whether or not to allow a read or a write a key
1256
value (e.g. `db1:user:1` for user `spencer`), a RoachNode would
1257
read the following permissions values:
1258
1259
```
1260
\0perm<db1:user:1>spencer
1261
\0perm<db1:user>spencer
1262
\0perm<db1>spencer
1263
\0perm<>spencer
1264
```
1265
1266
If any prefix in the hierarchy provides the required permission,
1267
the request is satisfied; otherwise, the request returns an
1268
error.
1269
1270
The priority for a user permission is used to order requests at
1271
Raft consensus ranges and for choosing an initial priority for
1272
distributed transactions. When scheduling operations at the Raft
1273
consensus range, all outstanding requests are ordered by key
1274
prefix and each assigned priorities according to key, user and
1275
arrival time. The next request is chosen probabilistically using
1276
priorities to weight the choice. Each key can have multiple
1277
priorities as they’re hierarchical (e.g. for /user/key, one
1278
priority for root ‘/’, and one for ‘/user/key’). The most general
1279
priority is used first. If two keys share the most general, then
1280
they’re compared with the next most general if applicable, and so on.
1281
1282
# Key-Value API
1283
1284
see the protobufs in [proto/](https://github.com/cockroachdb/cockroach/blob/master/proto),
1285
in particular [proto/api.proto](https://github.com/cockroachdb/cockroach/blob/master/proto/api.proto) and the comments within.
1286
1287
# Structured Data API
1288
1289
A preliminary design can be found in the [Go source documentation](http://godoc.org/github.com/cockroachdb/cockroach/structured).