Permalink
Apr 23, 2015
May 21, 2015
Apr 23, 2015
May 22, 2015
Apr 23, 2015
May 22, 2015
Apr 23, 2015
Apr 23, 2015
May 25, 2015
Apr 23, 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
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
May 29, 2015
Apr 23, 2015
May 28, 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
1271 lines (1056 sloc)
60.5 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
140
# Lock-Free Distributed Transactions
141
142
Cockroach provides distributed transactions without locks. Cockroach
143
transactions support two isolation levels:
144
145
- snapshot isolation (SI) and
146
- *serializable* snapshot isolation (SSI).
147
148
*SI* is simple to implement, highly performant, and correct for all but a
149
handful of anomalous conditions (e.g. write skew). *SSI* requires just a touch
150
more complexity, is still highly performant (less so with contention), and has
151
no anomalous conditions. Cockroach’s SSI implementation is based on ideas from
152
the literature and some possibly novel insights.
153
154
SSI is the default level, with SI provided for application developers
155
who are certain enough of their need for performance and the absence of
156
write skew conditions to consciously elect to use it. In a lightly
157
contended system, our implementation of SSI is just as performant as SI,
158
requiring no locking or additional writes. With contention, our
159
implementation of SSI still requires no locking, but will end up
160
aborting more transactions. Cockroach’s SI and SSI implementations
161
prevent starvation scenarios even for arbitrarily long transactions.
162
163
See the [Cahill paper](https://drive.google.com/file/d/0B9GCVTp_FHJIcEVyZVdDWEpYYXVVbFVDWElrYUV0NHFhU2Fv/edit?usp=sharing)
164
for one possible implementation of SSI. This is another [great paper](http://cs.yale.edu/homes/thomson/publications/calvin-sigmod12.pdf).
165
For a discussion of SSI implemented by preventing read-write conflicts
166
(in contrast to detecting them, called write-snapshot isolation), see
167
the [Yabandeh paper](https://drive.google.com/file/d/0B9GCVTp_FHJIMjJ2U2t6aGpHLTFUVHFnMTRUbnBwc2pLa1RN/edit?usp=sharing),
168
which is the source of much inspiration for Cockroach’s SSI.
169
170
Each Cockroach transaction is assigned a random priority and a
171
"candidate timestamp" at start. The candidate timestamp is the
172
provisional timestamp at which the transaction will commit, and is
173
chosen as the current clock time of the node coordinating the
174
transaction. This means that a transaction without conflicts will
175
usually commit with a timestamp that, in absolute time, precedes the
176
actual work done by that transaction.
177
178
In the course of coordinating a transaction between one or more
179
distributed nodes, the candidate timestamp may be increased, but will
180
never be decreased. The core difference between the two isolation levels
181
SI and SSI is that the former allows the transaction's candidate
182
timestamp to increase and the latter does not.
183
184
Timestamps are a combination of both a physical and a logical component
185
to support monotonic increments without degenerate cases causing
186
timestamps to diverge from wall clock time, following closely the
187
[*Hybrid Logical Clock
188
paper.*](http://www.cse.buffalo.edu/tech-reports/2014-04.pdf)
189
191
192
1. Start the transaction by writing a new entry to the system
193
transaction table (keys prefixed by *\0tx*) with state “PENDING”. In
194
parallel write an "intent" value for each datum being written as part
195
of the transaction. These are normal MVCC values, with the addition of
196
a special flag (i.e. “intent”) indicating that the value may be
198
the transaction id (unique and chosen at tx start time by client)
199
is stored with intent values. The tx id is used to refer to the
200
transaction table when there are conflicts and to make
201
tie-breaking decisions on ordering between identical timestamps.
204
selects the maximum from amongst all writes as the final commit
205
timestamp.
206
207
Each range maintains a small (i.e. latest 10s of read timestamps),
208
*in-memory* cache from key to the latest timestamp at which the
209
key(s) were read. This *latest-read-cache* is consulted on each
210
write. If the write’s candidate timestamp is earlier than the low
211
water mark on the cache itself (i.e. its last evicted timestamp)
212
or if the key being written has a read timestamp later than the
213
write’s candidate timestamp, this later timestamp value is
214
returned with the write. The cache’s entries are evicted oldest
215
timestamp first, updating low water mark as appropriate. If a new
216
range replica leader is elected, it sets the low water mark for
217
the cache to the current wall time + ε (ε = 99^th^ percentile
218
clock skew).
219
221
transaction table (keys prefixed by *\0tx*). The value of the
222
commit entry contains the candidate timestamp (increased as
223
necessary to accommodate any latest read timestamps). Note that
224
the transaction is considered fully committed at this point and
225
control may be returned to the client.
226
227
In the case of an SI transaction, a commit timestamp which was
228
increased to accommodate concurrent readers is perfectly
229
acceptable and the commit may continue. For SSI transactions,
230
however, a gap between candidate and commit timestamps
231
necessitates transaction restart (note: restart is different than
232
abort--see below).
233
234
After the transaction is committed, all written intents are upgraded
235
in parallel by removing the “intent” flag. The transaction is
236
considered fully committed before this step and does not wait for
237
it to return control to the transaction coordinator.
238
239
In the absence of conflicts, this is the end. Nothing else is necessary
240
to ensure the correctness of the system.
241
242
**Conflict Resolution**
243
244
Things get more interesting when a reader or writer encounters an intent
245
record or newly-committed value in a location that it needs to read or
246
write. This is a conflict, usually causing either of the transactions to
247
abort or restart depending on the type of conflict.
248
249
***Transaction restart:***
250
251
This is the usual (and more efficient) type of behaviour and is used
252
except when the transaction was aborted (for instance by another
253
transaction).
254
In effect, that reduces to two cases; the first being the one outlined
255
above: An SSI transaction that finds upon attempting to commit that
256
its commit timestamp has been pushed. The second case involves a transaction
257
actively encountering a conflict, that is, one of its readers or writers
258
encounter data that necessitate conflict resolution
259
(see transaction interactions below).
260
261
When a transaction restarts, it changes its priority and/or moves its
263
begins anew reusing the same tx id. The prior run of the transaction might
264
have written some write intents, which need to be deleted before the
265
transaction commits, so as to not be included as part of the transaction.
266
These stale write intent deletions are done during the reexecution of the
267
transaction, either implicitly, through writing new intents to
268
the same keys as part of the reexecution of the transaction, or explicitly,
269
by cleaning up stale intents that are not part of the reexecution of the
270
transaction. Since most transactions will end up writing to the same keys,
271
the explicit cleanup run just before committing the transaction is usually
272
a NOOP.
273
274
***Transaction abort:***
275
276
This is the case in which a transaction, upon reading its transaction
277
table entry, finds that it has been aborted. In this case, the
278
transaction can not reuse its intents; it returns control to the client
279
before cleaning them up (other readers and writers would clean up
280
dangling intents as they encounter them) but will make an effort to
281
clean up after itself. The next attempt (if applicable) then runs as a
285
286
There are several scenarios in which transactions interact:
287
288
- **Reader encounters write intent or value with newer timestamp far
289
enough in the future**: This is not a conflict. The reader is free
290
to proceed; after all, it will be reading an older version of the
291
value and so does not conflict. Recall that the write intent may
292
be committed with a later timestamp than its candidate; it will
293
never commit with an earlier one. **Side note**: if a SI transaction
294
reader finds an intent with a newer timestamp which the reader’s own
296
297
- **Reader encounters write intent or value with newer timestamp in the
298
near future:** In this case, we have to be careful. The newer
299
intent may, in absolute terms, have happened in our read's past if
300
the clock of the writer is ahead of the node serving the values.
301
In that case, we would need to take this value into account, but
302
we just don't know. Hence the transaction restarts, using instead
303
a future timestamp (but remembering a maximum timestamp used to
304
limit the uncertainty window to the maximum clock skew). In fact,
305
this is optimized further; see the details under "choosing a time
306
stamp" below.
307
308
- **Reader encounters write intent with older timestamp**: the reader
309
must follow the intent’s transaction id to the transaction table.
310
If the transaction has already been committed, then the reader can
311
just read the value. If the write transaction has not yet been
312
committed, then the reader has two options. If the write conflict
313
is from an SI transaction, the reader can *push that transaction's
314
commit timestamp into the future* (and consequently not have to
315
read it). This is simple to do: the reader just updates the
316
transaction’s commit timestamp to indicate that when/if the
317
transaction does commit, it should use a timestamp *at least* as
318
high. However, if the write conflict is from an SSI transaction,
319
the reader must compare priorities. If the reader has the higher priority,
320
it pushes the transaction’s commit timestamp (that
321
transaction will then notice its timestamp has been pushed, and
323
a new priority `max(new random priority, conflicting txn’s
325
327
If the other write intent has been written by a transaction with a lower
328
priority, the writer aborts the conflicting transaction. If the write
329
intent has a higher or equal priority the transaction retries, using as a new
330
priority *max(new random priority, conflicting txn’s priority - 1)*;
331
the retry occurs after a short, randomized backoff interval.
332
333
- **Writer encounters newer committed value**:
334
The committed value could also be an unresolved write intent made by a
335
transaction that has already committed. The transaction restarts. On restart,
336
the same priority is reused, but the candidate timestamp is moved forward
337
to the encountered value's timestamp.
338
339
**Transaction management**
340
341
Transactions are managed by the client proxy (or gateway in SQL Azure
342
parlance). Unlike in Spanner, writes are not buffered but are sent
343
directly to all implicated ranges. This allows the transaction to abort
344
quickly if it encounters a write conflict. The client proxy keeps track
345
of all written keys in order to resolve write intents asynchronously upon
346
transaction completion. If a transaction commits successfully, all intents
347
are upgraded to committed. In the event a transaction is aborted, all written
348
intents are deleted. The client proxy doesn’t guarantee it will resolve intents.
349
350
In the event the client proxy restarts before the pending transaction is
352
transaction table until aborted by another transaction. Transactions
353
heartbeat the transaction table every five seconds by default.
354
Transactions encountered by readers or writers with dangling intents
355
which haven’t been heartbeat within the required interval are aborted.
360
361
An exploration of retries with contention and abort times with abandoned
362
transaction is
363
[here](https://docs.google.com/document/d/1kBCu4sdGAnvLqpT-_2vaTbomNmX3_saayWEGYu1j7mQ/edit?usp=sharing).
364
365
**Transaction Table**
366
367
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`.
368
369
**Pros**
370
371
- No requirement for reliable code execution to prevent stalled 2PC
372
protocol.
373
- Readers never block with SI semantics; with SSI semantics, they may
374
abort.
375
- Lower latency than traditional 2PC commit protocol (w/o contention)
376
because second phase requires only a single write to the
377
transaction table instead of a synchronous round to all
378
transaction participants.
379
- Priorities avoid starvation for arbitrarily long transactions and
380
always pick a winner from between contending transactions (no
381
mutual aborts).
382
- Writes not buffered at client; writes fail fast.
383
- No read-locking overhead required for *serializable* SI (in contrast
384
to other SSI implementations).
385
- Well-chosen (i.e. less random) priorities can flexibly give
386
probabilistic guarantees on latency for arbitrary transactions
387
(for example: make OLTP transactions 10x less likely to abort than
388
low priority transactions, such as asynchronously scheduled jobs).
389
390
**Cons**
391
392
- Reads from non-leader replicas still require a ping to the leader to
393
update *latest-read-cache*.
394
- Abandoned transactions may block contending writers for up to the
395
heartbeat interval, though average wait is likely to be
396
considerably shorter (see [graph in link](https://docs.google.com/document/d/1kBCu4sdGAnvLqpT-_2vaTbomNmX3_saayWEGYu1j7mQ/edit?usp=sharing)).
397
This is likely considerably more performant than detecting and
398
restarting 2PC in order to release read and write locks.
399
- Behavior different than other SI implementations: no first writer
400
wins, and shorter transactions do not always finish quickly.
401
Element of surprise for OLTP systems may be a problematic factor.
402
- Aborts can decrease throughput in a contended system compared with
403
two phase locking. Aborts and retries increase read and write
404
traffic, increase latency and decrease throughput.
405
406
**Choosing a Timestamp**
407
408
A key challenge of reading data in a distributed system with clock skew
409
is choosing a timestamp guaranteed to be greater than the latest
410
timestamp of any committed transaction (in absolute time). No system can
411
claim consistency and fail to read already-committed data.
412
413
Time for a node is maintained by a hybrid logical clock (HLC) which combines
414
wall time and a logical time. The HLC time >= wall time and is affected by
415
transaction write requests from other nodes. Each transaction write request
416
from another node comes along with a timestamp. The timestamp is not only used
417
to version the data being written, but also updates the logical time on the
418
node. This is useful in guaranteeing that all data written to a node is at a
419
timestamp < HLC time
420
421
Accomplishing consistency for transactions (or just single operations)
422
accessing a single node is easy. The transaction uses the HLC time as the
423
timestamp which is guaranteed to be at a greater timestamp than all the
424
timestamped data on the node.
425
426
For multiple nodes, the timestamp of the node coordinating the
427
transaction `t` is used. In addition, a maximum timestamp `t+ε` is
428
supplied to provide an upper bound on timestamps for already-committed
429
data (`ε` is the maximum clock skew). As the transaction progresses, any
430
data read which have timestamps greater than `t` but less than `t+ε`
431
cause the transaction to abort and retry with the conflicting timestamp
432
t<sub>c</sub>, where t<sub>c</sub> \> t. The maximum timestamp `t+ε` remains
433
the same.
434
436
by uncertainty. Upon restarting, the transaction not only takes
437
into account t<sub>c<sub>, but the timestamp of the node at the time
438
of the uncertain read t<sub>node<sub>. The larger of those two timestamps
439
(more likely the latter): max(t<sub>c<sub>, t<sub>node<sub>) is used
440
to bump up the read timestamp. Additionally, the conflicting node is
441
marked as “certain”. This means that for future reads to that node
442
within the transaction, we can set `MaxTimestamp = Read Timestamp`.
443
Correctness follows from the fact that we know that at the time of the read,
444
there exists no version of any key on that node with a higher timestamp than
445
t<sub>node<sub>. Upon a restart caused by the node, if the transaction were to
446
encounter a key with a higher timestamp it would imply that the value
447
is written in the future in absolute time, and the transaction can move
448
forward reading an older version of the data (at the transactions timestamp).
449
This limits the time uncertainty restarts attributed to a node to <= 1. The
450
tradeoff is that we might pick a timestamp larger than the optimal one
451
(> highest conflicting timestamp), resulting in the possibility of a few
452
more conflicts.
453
454
We expect retries will be rare, but this assumption may need to be
455
revisited if retries become problematic. Note that this problem does not
456
apply to historical reads. An alternate approach which does not require
457
retries makes a round to all node participants in advance and
458
chooses the highest reported node wall time as the timestamp. However,
459
knowing which nodes will be accessed in advance is difficult and
460
potentially limiting. Cockroach could also potentially use a global
461
clock (Google did this with [Percolator](https://www.usenix.org/legacy/event/osdi10/tech/full_papers/Peng.pdf)), which would be
462
feasible for smaller, geographically-proximate clusters.
463
464
# Linearizability
465
466
First a word about [***Spanner***](http://research.google.com/archive/spanner.html).
467
By combining judicious use of wait intervals with accurate time signals,
468
Spanner provides a global ordering between any two non-overlapping transactions
469
(in absolute time) with \~14ms latencies. Put another way:
470
Spanner guarantees that if a transaction T<sub>1</sub> commits (in absolute time)
471
before another transaction T<sub>2</sub> starts, then T<sub>1</sub>'s assigned commit
472
timestamp is smaller than T<sub>2</sub>'s. Using atomic clocks and GPS receivers,
473
Spanner reduces their clock skew uncertainty to \< 10ms (`ε`). To make
474
good on the promised guarantee, transactions must take at least double
475
the clock skew uncertainty interval to commit (`2ε`). See [*this
476
article*](http://www.cs.cornell.edu/~ie53/publications/DC-col51-Sep13.pdf)
477
for a helpful overview of Spanner’s concurrency control.
478
479
Cockroach could make the same guarantees without specialized hardware,
480
at the expense of longer wait times. If servers in the cluster were
481
configured to work only with NTP, transaction wait times would likely to
482
be in excess of 150ms. For wide-area zones, this would be somewhat
483
mitigated by overlap from cross datacenter link latencies. If clocks
484
were made more accurate, the minimal limit for commit latencies would
485
improve.
486
487
However, let’s take a step back and evaluate whether Spanner’s external
488
consistency guarantee is worth the automatic commit wait. First, if the
489
commit wait is omitted completely, the system still yields a consistent
490
view of the map at an arbitrary timestamp. However with clock skew, it
491
would become possible for commit timestamps on non-overlapping but
492
causally related transactions to suffer temporal reverse. In other
493
words, the following scenario is possible for a client without global
494
ordering:
495
496
- Start transaction T<sub>1</sub> to modify value `x` with commit time *s<sub>1</sub>*
497
498
- On commit of T<sub>1</sub>, start T<sub>2</sub> to modify value `y` with commit time
499
\> s<sub>2</sub>
500
501
- Read `x` and `y` and discover that s<sub>1</sub> \> s<sub>2</sub> (**!**)
502
503
The external consistency which Spanner guarantees is referred to as
504
**linearizability**. It goes beyond serializability by preserving
505
information about the causality inherent in how external processes
506
interacted with the database. The strength of Spanner’s guarantee can be
507
formulated as follows: any two processes, with clock skew within
508
expected bounds, may independently record their wall times for the
509
completion of transaction T<sub>1</sub> (T<sub>1</sub><sup>end</sup>) and start of transaction
510
T<sub>2</sub> (T<sub>2</sub><sup>start</sup>) respectively, and if later
511
compared such that T<sub>1</sub><sup>end</sup> \< T<sub>2</sub><sup>start</sup>,
512
then commit timestamps s<sub>1</sub> \< s<sub>2</sub>.
513
This guarantee is broad enough to completely cover all cases of explicit
514
causality, in addition to covering any and all imaginable scenarios of implicit
515
causality.
516
517
Our contention is that causality is chiefly important from the
518
perspective of a single client or a chain of successive clients (*if a
519
tree falls in the forest and nobody hears…*). As such, Cockroach
520
provides two mechanisms to provide linearizability for the vast majority
521
of use cases without a mandatory transaction commit wait or an elaborate
522
system to minimize clock skew.
523
524
1. Clients provide the highest transaction commit timestamp with
525
> successive transactions. This allows node clocks from previous
526
> transactions to effectively participate in the formulation of the
527
> commit timestamp for the current transaction. This guarantees
528
> linearizability for transactions committed by this client.
529
>
530
> Newly launched clients wait at least 2 \* ε from process start
531
> time before beginning their first transaction. This preserves the
532
> same property even on client restart, and the wait will be
533
> mitigated by process initialization.
534
>
535
> All causally-related events within Cockroach maintain
536
> linearizability. Message queues, for example, guarantee that the
537
> receipt timestamp is greater than send timestamp, and that
538
> delivered messages may not be reaped until after the commit wait.
539
540
2. Committed transactions respond with a commit wait parameter which
541
> represents the remaining time in the nominal commit wait. This
542
> will typically be less than the full commit wait as the consensus
543
> write at the coordinator accounts for a portion of it.
544
>
545
> Clients taking any action outside of another Cockroach transaction
546
> (e.g. writing to another distributed system component) can either
547
> choose to wait the remaining interval before proceeding, or
548
> alternatively, pass the wait and/or commit timestamp to the
549
> execution of the outside action for its consideration. This pushes
550
> the burden of linearizability to clients, but is a useful tool in
551
> mitigating commit latencies if the clock skew is potentially
552
> large. This functionality can be used for ordering in the face of
553
> backchannel dependencies as mentioned in the
554
> [AugmentedTime](http://www.cse.buffalo.edu/~demirbas/publications/augmentedTime.pdf)
555
> paper.
556
557
Using these mechanisms in place of commit wait, Cockroach’s guarantee can be
558
formulated as follows: any process which signals the start of transaction
559
T<sub>2</sub> (T<sub>2</sub><sup>start</sup>) after the completion of
560
transaction T<sub>1</sub> (T<sub>1</sub><sup>end</sup>), will have commit
561
timestamps such thats<sub>1</sub> \< s<sub>2</sub>.
562
563
# Logical Map Content
564
565
Logically, the map contains a series of reserved system key / value
566
pairs covering accounting, range metadata, node accounting and
567
permissions before the actual key / value pairs for non-system data
568
(e.g. the actual meat of the map).
569
570
- `\0\0meta1` Range metadata for location of `\0\0meta2`.
571
- `\0\0meta1<key1>` Range metadata for location of `\0\0meta2<key1>`.
572
- ...
573
- `\0\0meta1<keyN>`: Range metadata for location of `\0\0meta2<keyN>`.
574
- `\0\0meta2`: Range metadata for location of first non-range metadata key.
575
- `\0\0meta2<key1>`: Range metadata for location of `<key1>`.
576
- ...
577
- `\0\0meta2<keyN>`: Range metadata for location of `<keyN>`.
578
- `\0acct<key0>`: Accounting for key prefix key0.
579
- ...
580
- `\0acct<keyN>`: Accounting for key prefix keyN.
581
- `\0node<node-address0>`: Accounting data for node 0.
582
- ...
583
- `\0node<node-addressN>`: Accounting data for node N.
584
- `\0perm<key0><user0>`: Permissions for user0 for key prefix key0.
585
- ...
586
- `\0perm<keyN><userN>`: Permissions for userN for key prefix keyN.
587
- `\0tree_root`: Range key for root of range-spanning tree.
588
- `\0tx<tx-id0>`: Transaction record for transaction 0.
589
- ...
590
- `\0tx<tx-idN>`: Transaction record for transaction N.
591
- `\0zone<key0>`: Zone information for key prefix key0.
592
- ...
593
- `\0zone<keyN>`: Zone information for key prefix keyN.
594
- `<>acctd<metric0>`: Accounting data for Metric 0 for empty key prefix.
595
- ...
596
- `<>acctd<metricN>`: Accounting data for Metric N for empty key prefix.
597
- `<key0>`: `<value0>` The first user data key.**
598
- ...
599
- `<keyN>`: `<valueN>` The last user data key.**
600
601
There are some additional system entries sprinkled amongst the
602
non-system keys. See the Key-Prefix Accounting section in this document
603
for further details.
604
605
# Node Storage
606
607
Nodes maintain a separate instance of RocksDB for each disk. Each
608
RocksDB instance hosts any number of ranges. RPCs arriving at a
609
RoachNode are multiplexed based on the disk name to the appropriate
610
RocksDB instance. A single instance per disk is used to avoid
611
contention. If every range maintained its own RocksDB, global management
612
of available cache memory would be impossible and writers for each range
613
would compete for non-contiguous writes to multiple RocksDB logs.
614
615
In addition to the key/value pairs of the range itself, various range
616
metadata is maintained.
617
618
- range-spanning tree node links
619
620
- participating replicas
621
622
- consensus metadata
623
624
- split/merge activity
625
626
A really good reference on tuning Linux installations with RocksDB is
627
[here](http://docs.basho.com/riak/latest/ops/advanced/backends/leveldb/).
628
629
# Range Metadata
630
631
The default approximate size of a range is 64M (2\^26 B). In order to
632
support 1P (2\^50 B) of logical data, metadata is needed for roughly
633
2\^(50 - 26) = 2\^24 ranges. A reasonable upper bound on range metadata
635
locations and 220 bytes for the range key itself*). 2\^24 ranges \* 2\^8
636
B would require roughly 4G (2\^32 B) to store--too much to duplicate
637
between machines. Our conclusion is that range metadata must be
638
distributed for large installations.
639
640
To distribute the range metadata and keep key lookups relatively fast,
641
we use two levels of indirection. All of the range metadata sorts first
642
in our key-value map. We accomplish this by prefixing range metadata
643
with two null characters (*\0\0*). The *meta1* or *meta2* suffixes are
644
additionally appended to distinguish between the first level and second
646
we first locate the range information for the lower bound of
647
`\0\0meta1<key1>`, and then use that range to locate the lower bound
648
of `\0\0meta2<key1>`. The range specified there will indicate the
649
range location of `<key1>` (refer to examples below). Using two levels
650
of indirection, **our map can address approximately 2\^62 B of data, or
651
roughly 4E** (*each metadata range addresses 2\^(26-8) = 2\^18 ranges;
652
with two levels of indirection, we can address 2\^(18 + 18) = 2\^36
653
ranges; each range addresses 2\^26 B; total is 2\^(36+26) B = 2\^62 B =
654
4E*).
655
656
Note: we append the end key of each range to meta[12] records because
657
the RocksDB iterator only supports a Seek() interface which acts as a
658
Ceil(). Using the start key of the range would cause Seek() to find the
659
key *after* the meta indexing record we’re looking for, which would
660
result in having to back the iterator up, an option which is both less
661
efficient and not available in all cases.
662
663
The following example shows the directory structure for a map with
664
three ranges worth of data. Ellipses indicate additional key/value pairs to
665
fill an entire range of data. Except for the fact that splitting ranges
666
requires updates to the range metadata with knowledge of the metadata layout,
667
the range metadata itself requires no special treatment or bootstrapping.
668
669
**Range 0** (located on servers `dcrama1:8000`, `dcrama2:8000`,
670
`dcrama3:8000`)
671
672
- `\0\0meta1\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
673
- `\0\0meta2<lastkey0>`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
674
- `\0\0meta2<lastkey1>`: `dcrama4:8000`, `dcrama5:8000`, `dcrama6:8000`
675
- `\0\0meta2\xff`: `dcrama7:8000`, `dcrama8:8000`, `dcrama9:8000`
676
- ...
677
- `<lastkey0>`: `<lastvalue0>`
678
679
**Range 1** (located on servers `dcrama4:8000`, `dcrama5:8000`,
680
`dcrama6:8000`)
681
682
- ...
683
- `<lastkey1>`: `<lastvalue1>`
684
685
**Range 2** (located on servers `dcrama7:8000`, `dcrama8:8000`,
686
`dcrama9:8000`)
687
688
- ...
689
- `<lastkey2>`: `<lastvalue2>`
690
691
Consider a simpler example of a map containing less than a single
692
range of data. In this case, all range metadata and all data are
693
located in the same range:
694
695
**Range 0** (located on servers `dcrama1:8000`, `dcrama2:8000`,
696
`dcrama3:8000`)*
697
698
- `\0\0meta1\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
699
- `\0\0meta2\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
700
- `<key0>`: `<value0>`
701
- `...`
702
703
Finally, a map large enough to need both levels of indirection would
704
look like (note that instead of showing range replicas, this
705
example is simplified to just show range indexes):
706
707
**Range 0**
708
709
- `\0\0meta1<lastkeyN-1>`: Range 0
710
- `\0\0meta1\xff`: Range 1
711
- `\0\0meta2<lastkey1>`: Range 1
712
- `\0\0meta2<lastkey2>`: Range 2
713
- `\0\0meta2<lastkey3>`: Range 3
714
- ...
715
- `\0\0meta2<lastkeyN-1>`: Range 262143
716
717
**Range 1**
718
719
- `\0\0meta2<lastkeyN>`: Range 262144
720
- `\0\0meta2<lastkeyN+1>`: Range 262145
721
- ...
722
- `\0\0meta2\xff`: Range 500,000
723
- ...
724
- `<lastkey1>`: `<lastvalue1>`
725
726
**Range 2**
727
728
- ...
729
- `<lastkey2>`: `<lastvalue2>`
730
731
**Range 3**
732
733
- ...
734
- `<lastkey3>`: `<lastvalue3>`
735
736
**Range 262144**
737
738
- ...
739
- `<lastkeyN>`: `<lastvalueN>`
740
741
**Range 262145**
742
743
- ...
744
- `<lastkeyN+1>`: `<lastvalueN+1>`
745
746
Note that the choice of range `262144` is just an approximation. The
747
actual number of ranges addressable via a single metadata range is
748
dependent on the size of the keys. If efforts are made to keep key sizes
749
small, the total number of addressable ranges would increase and vice
750
versa.
751
752
From the examples above it’s clear that key location lookups require at
753
most three reads to get the value for `<key>`:
754
755
1. lower bound of `\0\0meta1<key>`
756
2. lower bound of `\0\0meta2<key>`,
757
3. `<key>`.
758
759
For small maps, the entire lookup is satisfied in a single RPC to Range 0. Maps
760
containing less than 16T of data would require two lookups. Clients cache both
761
levels of range metadata, and we expect that data locality for individual
762
clients will be high. Clients may end up with stale cache entries. If on a
763
lookup, the range consulted does not match the client’s expectations, the
764
client evicts the stale entries and possibly does a new lookup.
765
766
# Range-Spanning Binary Tree
767
768
A crucial enhancement to the organization of range metadata is to
769
augment the bi-level range metadata lookup with a minimum spanning tree,
770
implemented as a left-leaning red-black tree over all ranges in the map.
771
This tree structure allows the system to start at any key prefix and
772
efficiently traverse an arbitrary key range with minimal RPC traffic,
773
minimal fan-in and fan-out, and with bounded time complexity equal to
774
`2*log N` steps, where `N` is the total number of ranges in the system.
775
776
Unlike the range metadata rows prefixed with `\0\0meta[1|2]`, the
777
metadata for the range-spanning tree (e.g. parent range and left / right
778
child ranges) is stored directly at the ranges as non-map metadata. The
779
metadata for each node of the tree (e.g. links to parent range, left
780
child range, and right child range) is stored with the range metadata.
781
In effect, the tree metadata is stored implicitly. In order to traverse
782
the tree, for example, you’d need to query each range in turn for its
783
metadata.
784
785
Any time a range is split or merged, both the bi-level range lookup
786
metadata and the per-range binary tree metadata are updated as part of
787
the same distributed transaction. The total number of nodes involved in
788
the update is bounded by 2 + log N (i.e. 2 updates for meta1 and
789
meta2, and up to log N updates to balance the range-spanning tree).
790
The range corresponding to the root node of the tree is stored in
792
793
As an example, consider the following set of nine ranges and their
794
associated range-spanning tree:
795
796
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`.
797
798
799
800
The range-spanning tree has many beneficial uses in Cockroach. It makes
801
the problem of efficiently aggregating accounting information of
802
potentially vast ranges of data tractable. Imagine a subrange of data
803
over which accounting is being kept. For example, the *photos* table in
804
a public photo sharing site. To efficiently keep track of data about the
805
table (e.g. total size, number of rows, etc.), messages can be passed
806
first up the tree and then down to the left until updates arrive at the
807
key prefix under which accounting is aggregated. This makes worst case
808
number of hops for an update to propagate into the accounting totals
809
2 \* log N. A 64T database will require 1M ranges, meaning 40 hops
810
worst case. In our experience, accounting tasks over vast ranges of data
811
are most often map/reduce jobs scheduled with coarse-grained
812
periodicity. By contrast, we expect Cockroach to maintain statistics
813
with sub 10s accuracy and with minimal cycles and minimal IOPs.
814
815
Another use for the range-spanning tree is to push accounting, zones and
816
permissions configurations to all ranges. In the case of zones and
817
permissions, this is an efficient way to pass updated configuration
818
information with exponential fan-out. When adding accounting
819
configurations (i.e. specifying a new key prefix to track), the
820
implicated ranges are transactionally scanned and zero-state accounting
821
information is computed as well. Deleting accounting configurations is
822
similar, except accounting records are deleted.
823
824
Last but *not* least, the range-spanning tree provides a convenient
825
mechanism for planning and executing parallel queries. These provide the
826
basis for
827
[Dremel](http://static.googleusercontent.com/media/research.google.com/en/us/pubs/archive/36632.pdf)-like
828
query execution trees and it’s easy to imagine supporting a subset of
829
SQL or even javascript-based user functions for complex data analysis
830
tasks.
831
832
# Raft - Consistency of Range Replicas
833
834
Each range is configured to consist of three or more replicas. The
835
replicas in a range maintain their own instance of a distributed
836
consensus algorithm. We use the [*Raft consensus
837
algorithm*](https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf)
838
as it is simpler to reason about and includes a reference implementation
839
covering important details. Every write to replicas is logged twice.
840
Once to RocksDB’s internal log and once to levedb itself as part of the
841
Raft consensus log.
842
[ePaxos](https://www.cs.cmu.edu/~dga/papers/epaxos-sosp2013.pdf) has
843
promising performance characteristics for WAN-distributed replicas, but
844
it does not guarantee a consistent ordering between replicas.
845
846
Raft elects a relatively long-lived leader which must be involved to
847
propose writes. It heartbeats followers periodically to keep their logs
848
replicated. In the absence of heartbeats, followers become candidates
849
after randomized election timeouts and proceed to hold new leader
850
elections. Cockroach weights random timeouts such that the replicas with
851
shorter round trip times to peers are more likely to hold elections
852
first. Although only the leader can propose a new write, and as such
853
must be involved in any write to the consensus log, any replica can
854
service reads if the read is for a timestamp which the replica knows is
855
safe based on the last committed consensus write and the state of any
856
pending transactions.
857
858
Only the leader can propose a new write, but Cockroach accepts writes at
859
any replica. The replica merely forwards the write to the leader.
860
Instead of resending the write, the leader has only to acknowledge the
861
write to the forwarding replica using a log sequence number, as though
862
it were proposing it in the first place. The other replicas receive the
863
full write as though the leader were the originator.
864
865
Having a stable leader provides the choice of replica to handle
866
range-specific maintenance and processing tasks, such as delivering
867
pending message queues, handling splits and merges, rebalancing, etc.
868
869
# Splitting / Merging Ranges
870
871
RoachNodes split or merge ranges based on whether they exceed maximum or
872
minimum thresholds for capacity or load. Ranges exceeding maximums for
873
either capacity or load are split; ranges below minimums for *both*
874
capacity and load are merged.
875
876
Ranges maintain the same accounting statistics as accounting key
877
prefixes. These boil down to a time series of data points with minute
878
granularity. Everything from number of bytes to read/write queue sizes.
879
Arbitrary distillations of the accounting stats can be determined as the
880
basis for splitting / merging. Two sensical metrics for use with
881
split/merge are range size in bytes and IOps. A good metric for
882
rebalancing a replica from one node to another would be total read/write
883
queue wait times. These metrics are gossipped, with each range / node
884
passing along relevant metrics if they’re in the bottom or top of the
885
range it’s aware of.
886
887
A range finding itself exceeding either capacity or load threshold
888
splits. To this end, the range leader computes an appropriate split key
889
candidate and issues the split through Raft. In contrast to splitting,
890
merging requires a range to be below the minimum threshold for both
891
capacity *and* load. A range being merged chooses the smaller of the
892
ranges immediately preceding and succeeding it.
893
894
Splitting, merging, rebalancing and recovering all follow the same basic
895
algorithm for moving data between roach nodes. New target replicas are
896
created and added to the replica set of source range. Then each new
897
replica is brought up to date by either replaying the log in full or
898
copying a snapshot of the source replica data and then replaying the log
899
from the timestamp of the snapshot to catch up fully. Once the new
900
replicas are fully up to date, the range metadata is updated and old,
901
source replica(s) deleted if applicable.
902
903
**Coordinator** (leader replica)
904
908
only after being completed locally, are moved to new target replicas.
909
else if merging
910
Choose new replicas on same servers as target range replicas;
911
add to replica set.
912
else if rebalancing || recovering
913
Choose new replica(s) on least loaded servers; add to replica set.
914
```
915
916
**New Replica**
917
918
*Bring replica up to date:*
919
920
```
921
if all info can be read from replicated log
922
copy replicated log
923
else
924
snapshot source replica
925
send successive ReadRange requests to source replica
926
referencing snapshot
927
928
if merging
929
combine ranges on all replicas
930
else if rebalancing || recovering
931
remove old range replica(s)
932
```
933
934
RoachNodes split ranges when the total data in a range exceeds a
935
configurable maximum threshold. Similarly, ranges are merged when the
936
total data falls below a configurable minimum threshold.
937
938
**TBD: flesh this out**.
939
940
Ranges are rebalanced if a node determines its load or capacity is one
941
of the worst in the cluster based on gossipped load stats. A node with
942
spare capacity is chosen in the same datacenter and a special-case split
943
is done which simply duplicates the data 1:1 and resets the range
944
configuration metadata.
945
946
# Message Queues
947
948
Each range maintains an array of incoming message queues, referred to
949
here as **inboxes**. Additionally, each range maintains and *processes*
950
an array of outgoing message queues, referred to here as **outboxes**.
951
Both inboxes and outboxes are assigned to keys; messages can be sent or
952
received on behalf of any key. Inboxes and outboxes can contain any
953
number of pending messages.
954
955
Messages can be *deliverable*, or *executable.*
956
957
Deliverable messages are defined by Value objects - simple byte arrays -
958
that are delivered to a key’s inbox, awaiting collection by a client
959
invoking the ReapQueue operation. These are typically used by client
960
applications wishing to be notified of changes to an entry for further
961
processing, such as expensive offline operations like sending emails,
962
SMSs, etc.
963
964
Executable messages are *outgoing-only*, and are instances of
965
PutRequest,IncrementRequest, DeleteRequest, DeleteRangeRequest
967
executed when encountered. These are primarily useful when updates that
968
are nominally part of a transaction can tolerate asynchronous execution
969
(e.g. eventual consistency), and are otherwise too busy or numerous to
970
make including them in the original [distributed] transaction efficient.
971
Examples may include updates to the accounting for successive key
972
prefixes (potentially busy) or updates to a full-text index (potentially
973
numerous).
974
975
These two types of messages are enqueued in different outboxes too - see
976
key formats below.
977
978
At commit time, the range processing the transaction places messages
979
into a shared outbox located at the start of the range metadata. This is
980
effectively free as it’s part of the same consensus write for the range
981
as the COMMIT record. Outgoing messages are processed asynchronously by
982
the range. To make processing easy, all outboxes are co-located at the
983
start of the range. To make lookup easy, all inboxes are located
984
immediately after the recipient key. The leader replica of a range is
985
responsible for processing message queues.
986
987
A dispatcher polls a given range’s *deliverable message outbox*
988
periodically (configurable), and delivers those messages to the target
989
key’s inbox. The dispatcher is also woken up whenever a new message is
990
added to the outbox. A separate executor also polls the range’s
991
*executable message outbox* periodically as well (again, configurable),
993
new message is added to the outbox.
994
995
Formats follow in the table below. Notice that inbox messages for a
996
given key sort by the `<outbox-timestamp>`. This doesn’t provide a
997
precise ordering, but it does allow clients to scan messages in an
998
approximate ordering of when they were originally lodged with senders.
999
NTP offers walltime deltas to within 100s of milliseconds. The
1000
`<sender-range-key>` suffix provides uniqueness.
1001
1002
**Outbox**
1003
`<sender-range-key>deliverable-outbox:<recipient-key><outbox-timestamp>`
1004
`<sender-range-key>executable-outbox:<recipient-key><outbox-timestamp>`
1005
1006
**Inbox**
1007
`<recipient-key>inbox:<outbox-timestamp><sender-range-key>`
1008
1009
Messages are processed and then deleted as part of a single distributed
1010
transaction. The message will be executed or delivered exactly once,
1011
regardless of failures at either sender or receiver.
1012
1013
Delivered messages may be read by clients via the ReapQueue operation.
1014
This operation may only be used as part of a transaction. Clients should
1015
commit only after having processed the message. If the transaction is
1016
committed, scanned messages are automatically deleted. The operation
1017
name was chosen to reflect its mutating side effect. Deletion of read
1018
messages is mandatory because senders deliver messages asynchronously
1019
and a delay could cause insertion of messages at arbitrary points in the
1020
inbox queue. If clients require persistence, they should re-save read
1021
messages manually; the ReapQueue operation can be incorporated into
1022
normal transactional updates.
1023
1024
# Node Allocation (via Gossip)
1025
1026
New nodes must be allocated when a range is split. Instead of requiring
1027
every RoachNode to know about the status of all or even a large number
1028
of peer nodes --or-- alternatively requiring a specialized curator or
1029
master with sufficiently global knowledge, we use a gossip protocol to
1030
efficiently communicate only interesting information between all of the
1031
nodes in the cluster. What’s interesting information? One example would
1032
be whether a particular node has a lot of spare capacity. Each node,
1033
when gossiping, compares each topic of gossip to its own state. If its
1034
own state is somehow “more interesting” than the least interesting item
1035
in the topic it’s seen recently, it includes its own state as part of
1036
the next gossip session with a peer node. In this way, a node with
1037
capacity sufficiently in excess of the mean quickly becomes discovered
1038
by the entire cluster. To avoid piling onto outliers, nodes from the
1039
high capacity set are selected at random for allocation.
1040
1041
The gossip protocol itself contains two primary components:
1042
1043
- **Peer Selection**: each node maintains up to N peers with which it
1044
regularly communicates. It selects peers with an eye towards
1045
maximizing fanout. A peer node which itself communicates with an
1046
array of otherwise unknown nodes will be selected over one which
1047
communicates with a set containing significant overlap. Each time
1048
gossip is initiated, each nodes’ set of peers is exchanged. Each
1049
node is then free to incorporate the other’s peers as it sees fit.
1050
To avoid any node suffering from excess incoming requests, a node
1051
may refuse to answer a gossip exchange. Each node is biased
1052
towards answering requests from nodes without significant overlap
1053
and refusing requests otherwise.
1054
1055
Peers are efficiently selected using a heuristic as described in
1056
[Agarwal & Trachtenberg (2006)](https://drive.google.com/file/d/0B9GCVTp_FHJISmFRTThkOEZSM1U/edit?usp=sharing).
1057
1058
**TBD**: how to avoid partitions? Need to work out a simulation of
1059
the protocol to tune the behavior and see empirically how well it
1060
works.
1061
1062
- **Gossip Selection**: what to communicate. Gossip is divided into
1063
topics. Load characteristics (capacity per disk, cpu load, and
1064
state [e.g. draining, ok, failure]) are used to drive node
1065
allocation. Range statistics (range read/write load, missing
1066
replicas, unavailable ranges) and network topology (inter-rack
1067
bandwidth/latency, inter-datacenter bandwidth/latency, subnet
1068
outages) are used for determining when to split ranges, when to
1069
recover replicas vs. wait for network connectivity, and for
1070
debugging / sysops. In all cases, a set of minimums and a set of
1071
maximums is propagated; each node applies its own view of the
1072
world to augment the values. Each minimum and maximum value is
1073
tagged with the reporting node and other accompanying contextual
1074
information. Each topic of gossip has its own protobuf to hold the
1075
structured data. The number of items of gossip in each topic is
1076
limited by a configurable bound.
1077
1078
For efficiency, nodes assign each new item of gossip a sequence
1079
number and keep track of the highest sequence number each peer
1080
node has seen. Each round of gossip communicates only the delta
1081
containing new items.
1082
1083
# Node Accounting
1084
1085
The gossip protocol discussed in the previous section is useful to
1086
quickly communicate fragments of important information in a
1087
decentralized manner. However, complete accounting for each node is also
1088
stored to a central location, available to any dashboard process. This
1089
is done using the map itself. Each node periodically writes its state to
1090
the map with keys prefixed by `\0node`, similar to the first level of
1091
range metadata, but with an ‘`node`’ suffix. Each value is a protobuf
1092
containing the full complement of node statistics--everything
1093
communicated normally via the gossip protocol plus other useful, but
1094
non-critical data.
1095
1096
The range containing the first key in the node accounting table is
1097
responsible for gossiping the total count of nodes. This total count is
1098
used by the gossip network to most efficiently organize itself. In
1099
particular, the maximum number of hops for gossipped information to take
1100
before reaching a node is given by `ceil(log(node count) / log(max
1101
fanout)) + 1`.
1102
1103
# Key-prefix Accounting, Zones & Permissions
1104
1105
Arbitrarily fine-grained accounting and permissions are specified via
1106
key prefixes. Key prefixes can overlap, as is necessary for capturing
1107
hierarchical relationships. For illustrative purposes, let’s say keys
1108
specifying rows in a set of databases have the following format:
1109
1110
`<db>:<table>:<primary-key>[:<secondary-key>]`
1111
1112
In this case, we might collect accounting or specify permissions with
1113
key prefixes:
1114
1115
`db1`, `db1:user`, `db1:order`,
1116
1117
Accounting is kept for the entire map by default.
1118
1119
## Accounting
1120
to keep accounting for a range defined by a key prefix, an entry is created in
1121
the accounting system table. The format of accounting table keys is:
1122
1123
`\0acct<key-prefix>`
1124
1125
In practice, we assume each RoachNode capable of caching the
1126
entire accounting table as it is likely to be relatively small.
1127
1128
Accounting is kept for key prefix ranges with eventual consistency
1129
for efficiency. Updates to accounting values propagate through the
1130
system using the message queue facility if the accounting keys do
1131
not reside on the same range as ongoing activity (true for all but
1132
the smallest systems). There are two types of values which
1133
comprise accounting: counts and occurrences, for lack of better
1134
terms. Counts describe system state, such as the total number of
1135
bytes, rows, etc. Occurrences include transient performance and
1136
load metrics. Both types of accounting are captured as time series
1137
with minute granularity. The length of time accounting metrics are
1138
kept is configurable. Below are examples of each type of
1139
accounting value.
1140
1141
**System State Counters/Performance**
1142
1143
- Count of items (e.g. rows)
1144
- Total bytes
1145
- Total key bytes
1146
- Total value length
1147
- Queued message count
1148
- Queued message total bytes
1149
- Count of values \< 16B
1150
- Count of values \< 64B
1151
- Count of values \< 256B
1152
- Count of values \< 1K
1153
- Count of values \< 4K
1154
- Count of values \< 16K
1155
- Count of values \< 64K
1156
- Count of values \< 256K
1157
- Count of values \< 1M
1158
- Count of values \> 1M
1159
- Total bytes of accounting
1160
1161
1162
**Load Occurences**
1163
1164
Get op count
1165
Get total MB
1166
Put op count
1167
Put total MB
1168
Delete op count
1169
Delete total MB
1170
Delete range op count
1171
Delete range total MB
1172
Scan op count
1173
Scan op MB
1174
Split count
1175
Merge count
1176
1177
Because accounting information is kept as time series and over many
1178
possible metrics of interest, the data can become numerous. Accounting
1179
data are stored in the map near the key prefix described, in order to
1180
distribute load (for both aggregation and storage).
1181
1182
Accounting keys for system state have the form:
1183
`<key-prefix>|acctd<metric-name>*`. Notice the leading ‘pipe’
1184
character. It’s meant to sort the root level account AFTER any other
1185
system tables. They must increment the same underlying values as they
1186
are permanent counts, and not transient activity. Logic at the
1187
RoachNode takes care of snapshotting the value into an appropriately
1188
suffixed (e.g. with timestamp hour) multi-value time series entry.
1189
1190
Keys for perf/load metrics:
1191
`<key-prefix>acctd<metric-name><hourly-timestamp>`.
1192
1193
`<hourly-timestamp>`-suffixed accounting entries are multi-valued,
1194
containing a varint64 entry for each minute with activity during the
1195
specified hour.
1196
1197
To efficiently keep accounting over large key ranges, the task of
1198
aggregation must be distributed. If activity occurs within the same
1199
range as the key prefix for accounting, the updates are made as part
1200
of the consensus write. If the ranges differ, then a message is sent
1201
to the parent range to increment the accounting. If upon receiving the
1202
message, the parent range also does not include the key prefix, it in
1203
turn forwards it to its parent or left child in the balanced binary
1204
tree which is maintained to describe the range hierarchy. This limits
1205
the number of messages before an update is visible at the root to `2*log N`,
1206
where `N` is the number of ranges in the key prefix.
1207
1208
## Zones
1209
zones are stored in the map with keys prefixed by
1210
`\0zone` followed by the key prefix to which the zone
1211
configuration applies. Zone values specify a protobuf containing
1212
the datacenters from which replicas for ranges which fall under
1213
the zone must be chosen.
1214
1215
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`.
1216
1217
If zones are modified in situ, each RoachNode verifies the
1218
existing zones for its ranges against the zone configuration. If
1219
it discovers differences, it reconfigures ranges in the same way
1220
that it rebalances away from busy nodes, via special-case 1:1
1221
split to a duplicate range comprising the new configuration.
1222
1223
### Permissions
1224
permissions are stored in the map with keys prefixed by *\0perm* followed by
1225
the key prefix and user to which the specified permissions apply. The format of
1226
permissions keys is:
1227
1228
`\0perm<key-prefix><user>`
1229
1230
Permission values are a protobuf containing the permission details;
1231
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`.
1232
1233
A default system root permission is assumed for the entire map
1234
with an empty key prefix and read/write as true.
1235
1236
When determining whether or not to allow a read or a write a key
1237
value (e.g. `db1:user:1` for user `spencer`), a RoachNode would
1238
read the following permissions values:
1239
1240
```
1241
\0perm<db1:user:1>spencer
1242
\0perm<db1:user>spencer
1243
\0perm<db1>spencer
1244
\0perm<>spencer
1245
```
1246
1247
If any prefix in the hierarchy provides the required permission,
1248
the request is satisfied; otherwise, the request returns an
1249
error.
1250
1251
The priority for a user permission is used to order requests at
1252
Raft consensus ranges and for choosing an initial priority for
1253
distributed transactions. When scheduling operations at the Raft
1254
consensus range, all outstanding requests are ordered by key
1255
prefix and each assigned priorities according to key, user and
1256
arrival time. The next request is chosen probabilistically using
1257
priorities to weight the choice. Each key can have multiple
1258
priorities as they’re hierarchical (e.g. for /user/key, one
1259
priority for root ‘/’, and one for ‘/user/key’). The most general
1260
priority is used first. If two keys share the most general, then
1261
they’re compared with the next most general if applicable, and so on.
1262
1263
# Key-Value API
1264
1265
see the protobufs in [proto/](https://github.com/cockroachdb/cockroach/blob/master/proto),
1266
in particular [proto/api.proto](https://github.com/cockroachdb/cockroach/blob/master/proto/api.proto) and the comments within.
1267
1268
# Structured Data API
1269
1270
A preliminary design can be found in the [Go source documentation](http://godoc.org/github.com/cockroachdb/cockroach/structured).