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
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
May 28, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
May 28, 2015
Apr 23, 2015
May 28, 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
1269 lines (1054 sloc)
60.3 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
346
If a transaction is completed successfully, all intents are upgraded to
347
committed. In the event a transaction is aborted, all written intents
349
351
In the event the client proxy restarts before the pending transaction is
353
transaction table until aborted by another transaction. Transactions
354
heartbeat the transaction table every five seconds by default.
355
Transactions encountered by readers or writers with dangling intents
356
which haven’t been heartbeat within the required interval are aborted.
357
In the event the proxy restarts after a transaction commits but before
358
the post processing is complete, the dangling intents are upgraded
359
when encountered by future readers and writers and the system does
360
not depend on their timely post-processing for correctness.
361
362
An exploration of retries with contention and abort times with abandoned
363
transaction is
364
[here](https://docs.google.com/document/d/1kBCu4sdGAnvLqpT-_2vaTbomNmX3_saayWEGYu1j7mQ/edit?usp=sharing).
365
366
**Transaction Table**
367
368
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`.
369
370
**Pros**
371
372
- No requirement for reliable code execution to prevent stalled 2PC
373
protocol.
374
- Readers never block with SI semantics; with SSI semantics, they may
375
abort.
376
- Lower latency than traditional 2PC commit protocol (w/o contention)
377
because second phase requires only a single write to the
378
transaction table instead of a synchronous round to all
379
transaction participants.
380
- Priorities avoid starvation for arbitrarily long transactions and
381
always pick a winner from between contending transactions (no
382
mutual aborts).
383
- Writes not buffered at client; writes fail fast.
384
- No read-locking overhead required for *serializable* SI (in contrast
385
to other SSI implementations).
386
- Well-chosen (i.e. less random) priorities can flexibly give
387
probabilistic guarantees on latency for arbitrary transactions
388
(for example: make OLTP transactions 10x less likely to abort than
389
low priority transactions, such as asynchronously scheduled jobs).
390
391
**Cons**
392
393
- Reads from non-leader replicas still require a ping to the leader to
394
update *latest-read-cache*.
395
- Abandoned transactions may block contending writers for up to the
396
heartbeat interval, though average wait is likely to be
397
considerably shorter (see [graph in link](https://docs.google.com/document/d/1kBCu4sdGAnvLqpT-_2vaTbomNmX3_saayWEGYu1j7mQ/edit?usp=sharing)).
398
This is likely considerably more performant than detecting and
399
restarting 2PC in order to release read and write locks.
400
- Behavior different than other SI implementations: no first writer
401
wins, and shorter transactions do not always finish quickly.
402
Element of surprise for OLTP systems may be a problematic factor.
403
- Aborts can decrease throughput in a contended system compared with
404
two phase locking. Aborts and retries increase read and write
405
traffic, increase latency and decrease throughput.
406
407
**Choosing a Timestamp**
408
409
A key challenge of reading data in a distributed system with clock skew
410
is choosing a timestamp guaranteed to be greater than the latest
411
timestamp of any committed transaction (in absolute time). No system can
412
claim consistency and fail to read already-committed data.
413
414
Accomplishing this for transactions (or just single operations)
415
accessing a single node is easy. The transaction supplies 0 for
416
timestamp, indicating that the node should use its current time (time
417
for a node is kept using a hybrid clock which combines wall time and a
418
logical time). This guarantees data already committed to that node have
419
earlier timestamps.
420
421
For multiple nodes, the timestamp of the node coordinating the
422
transaction `t` is used. In addition, a maximum timestamp `t+ε` is
423
supplied to provide an upper bound on timestamps for already-committed
424
data (`ε` is the maximum clock skew). As the transaction progresses, any
425
data read which have timestamps greater than `t` but less than `t+ε`
426
cause the transaction to abort and retry with the conflicting timestamp
427
t<sub>c</sub>, where t<sub>c</sub> \> t. The maximum timestamp `t+ε` remains
428
the same.
429
430
We apply another optimization to avoid further restarts caused
431
by uncertainty. Upon restarting, the transaction not only takes
432
into account t<sub>c<sub>, but the timestamp of the node at the time
433
of the uncertain read t<>node<>. The larger of those two timestamps:
434
max(t<sub>c<sub>, t<sub>node<sub>) (more likely the latter) is used
435
to bump up the read timestamp. The use of t<sub>node<sub> guarantees that the
436
time spent retrying because of reading recently committed data has an
437
upper bound of `ε`.
438
439
Additionally, the conflicting node is marked as “certain”. This means that
440
for future reads to that node within the transaction, we can set
441
`MaxTimestamp = Read Timestamp`. Correctness follows from the fact that we
442
know that at the time of the read, there exists no version of any key on
443
that node with a higher timestamp that was updated in the past in absolute
444
time. A higher timestamp is indicative that the update happened in the future
445
in absolute time, and the transaction will move forward reading an older
446
version of the data (at the transactions timestamp). (This seems wrong; it only
447
guarantees that a transaction coordinated from this node cannot have a greater
448
timestamp. You can still have transactions coordinated from other nodes with
449
greater timestamps picked in the absolute past time that committed data to this
450
node.)
451
452
We expect retries will be rare, but this assumption may need to be
453
revisited if retries become problematic. Note that this problem does not
454
apply to historical reads. An alternate approach which does not require
455
retries makes a round to all node participants in advance and
456
chooses the highest reported node wall time as the timestamp. However,
457
knowing which nodes will be accessed in advance is difficult and
458
potentially limiting. Cockroach could also potentially use a global
459
clock (Google did this with [Percolator](https://www.usenix.org/legacy/event/osdi10/tech/full_papers/Peng.pdf)), which would be
460
feasible for smaller, geographically-proximate clusters.
461
462
# Linearizability
463
464
First a word about [***Spanner***](http://research.google.com/archive/spanner.html).
465
By combining judicious use of wait intervals with accurate time signals,
466
Spanner provides a global ordering between any two non-overlapping transactions
467
(in absolute time) with \~14ms latencies. Put another way:
468
Spanner guarantees that if a transaction T<sub>1</sub> commits (in absolute time)
469
before another transaction T<sub>2</sub> starts, then T<sub>1</sub>'s assigned commit
470
timestamp is smaller than T<sub>2</sub>'s. Using atomic clocks and GPS receivers,
471
Spanner reduces their clock skew uncertainty to \< 10ms (`ε`). To make
472
good on the promised guarantee, transactions must take at least double
473
the clock skew uncertainty interval to commit (`2ε`). See [*this
474
article*](http://www.cs.cornell.edu/~ie53/publications/DC-col51-Sep13.pdf)
475
for a helpful overview of Spanner’s concurrency control.
476
477
Cockroach could make the same guarantees without specialized hardware,
478
at the expense of longer wait times. If servers in the cluster were
479
configured to work only with NTP, transaction wait times would likely to
480
be in excess of 150ms. For wide-area zones, this would be somewhat
481
mitigated by overlap from cross datacenter link latencies. If clocks
482
were made more accurate, the minimal limit for commit latencies would
483
improve.
484
485
However, let’s take a step back and evaluate whether Spanner’s external
486
consistency guarantee is worth the automatic commit wait. First, if the
487
commit wait is omitted completely, the system still yields a consistent
488
view of the map at an arbitrary timestamp. However with clock skew, it
489
would become possible for commit timestamps on non-overlapping but
490
causally related transactions to suffer temporal reverse. In other
491
words, the following scenario is possible for a client without global
492
ordering:
493
494
- Start transaction T<sub>1</sub> to modify value `x` with commit time *s<sub>1</sub>*
495
496
- On commit of T<sub>1</sub>, start T<sub>2</sub> to modify value `y` with commit time
497
\> s<sub>2</sub>
498
499
- Read `x` and `y` and discover that s<sub>1</sub> \> s<sub>2</sub> (**!**)
500
501
The external consistency which Spanner guarantees is referred to as
502
**linearizability**. It goes beyond serializability by preserving
503
information about the causality inherent in how external processes
504
interacted with the database. The strength of Spanner’s guarantee can be
505
formulated as follows: any two processes, with clock skew within
506
expected bounds, may independently record their wall times for the
507
completion of transaction T<sub>1</sub> (T<sub>1</sub><sup>end</sup>) and start of transaction
508
T<sub>2</sub> (T<sub>2</sub><sup>start</sup>) respectively, and if later
509
compared such that T<sub>1</sub><sup>end</sup> \< T<sub>2</sub><sup>start</sup>,
510
then commit timestamps s<sub>1</sub> \< s<sub>2</sub>.
511
This guarantee is broad enough to completely cover all cases of explicit
512
causality, in addition to covering any and all imaginable scenarios of implicit
513
causality.
514
515
Our contention is that causality is chiefly important from the
516
perspective of a single client or a chain of successive clients (*if a
517
tree falls in the forest and nobody hears…*). As such, Cockroach
518
provides two mechanisms to provide linearizability for the vast majority
519
of use cases without a mandatory transaction commit wait or an elaborate
520
system to minimize clock skew.
521
522
1. Clients provide the highest transaction commit timestamp with
523
> successive transactions. This allows node clocks from previous
524
> transactions to effectively participate in the formulation of the
525
> commit timestamp for the current transaction. This guarantees
526
> linearizability for transactions committed by this client.
527
>
528
> Newly launched clients wait at least 2 \* ε from process start
529
> time before beginning their first transaction. This preserves the
530
> same property even on client restart, and the wait will be
531
> mitigated by process initialization.
532
>
533
> All causally-related events within Cockroach maintain
534
> linearizability. Message queues, for example, guarantee that the
535
> receipt timestamp is greater than send timestamp, and that
536
> delivered messages may not be reaped until after the commit wait.
537
538
2. Committed transactions respond with a commit wait parameter which
539
> represents the remaining time in the nominal commit wait. This
540
> will typically be less than the full commit wait as the consensus
541
> write at the coordinator accounts for a portion of it.
542
>
543
> Clients taking any action outside of another Cockroach transaction
544
> (e.g. writing to another distributed system component) can either
545
> choose to wait the remaining interval before proceeding, or
546
> alternatively, pass the wait and/or commit timestamp to the
547
> execution of the outside action for its consideration. This pushes
548
> the burden of linearizability to clients, but is a useful tool in
549
> mitigating commit latencies if the clock skew is potentially
550
> large. This functionality can be used for ordering in the face of
551
> backchannel dependencies as mentioned in the
552
> [AugmentedTime](http://www.cse.buffalo.edu/~demirbas/publications/augmentedTime.pdf)
553
> paper.
554
555
Using these mechanisms in place of commit wait, Cockroach’s guarantee can be
556
formulated as follows: any process which signals the start of transaction
557
T<sub>2</sub> (T<sub>2</sub><sup>start</sup>) after the completion of
558
transaction T<sub>1</sub> (T<sub>1</sub><sup>end</sup>), will have commit
559
timestamps such thats<sub>1</sub> \< s<sub>2</sub>.
560
561
# Logical Map Content
562
563
Logically, the map contains a series of reserved system key / value
564
pairs covering accounting, range metadata, node accounting and
565
permissions before the actual key / value pairs for non-system data
566
(e.g. the actual meat of the map).
567
568
- `\0\0meta1` Range metadata for location of `\0\0meta2`.
569
- `\0\0meta1<key1>` Range metadata for location of `\0\0meta2<key1>`.
570
- ...
571
- `\0\0meta1<keyN>`: Range metadata for location of `\0\0meta2<keyN>`.
572
- `\0\0meta2`: Range metadata for location of first non-range metadata key.
573
- `\0\0meta2<key1>`: Range metadata for location of `<key1>`.
574
- ...
575
- `\0\0meta2<keyN>`: Range metadata for location of `<keyN>`.
576
- `\0acct<key0>`: Accounting for key prefix key0.
577
- ...
578
- `\0acct<keyN>`: Accounting for key prefix keyN.
579
- `\0node<node-address0>`: Accounting data for node 0.
580
- ...
581
- `\0node<node-addressN>`: Accounting data for node N.
582
- `\0perm<key0><user0>`: Permissions for user0 for key prefix key0.
583
- ...
584
- `\0perm<keyN><userN>`: Permissions for userN for key prefix keyN.
585
- `\0tree_root`: Range key for root of range-spanning tree.
586
- `\0tx<tx-id0>`: Transaction record for transaction 0.
587
- ...
588
- `\0tx<tx-idN>`: Transaction record for transaction N.
589
- `\0zone<key0>`: Zone information for key prefix key0.
590
- ...
591
- `\0zone<keyN>`: Zone information for key prefix keyN.
592
- `<>acctd<metric0>`: Accounting data for Metric 0 for empty key prefix.
593
- ...
594
- `<>acctd<metricN>`: Accounting data for Metric N for empty key prefix.
595
- `<key0>`: `<value0>` The first user data key.**
596
- ...
597
- `<keyN>`: `<valueN>` The last user data key.**
598
599
There are some additional system entries sprinkled amongst the
600
non-system keys. See the Key-Prefix Accounting section in this document
601
for further details.
602
603
# Node Storage
604
605
Nodes maintain a separate instance of RocksDB for each disk. Each
606
RocksDB instance hosts any number of ranges. RPCs arriving at a
607
RoachNode are multiplexed based on the disk name to the appropriate
608
RocksDB instance. A single instance per disk is used to avoid
609
contention. If every range maintained its own RocksDB, global management
610
of available cache memory would be impossible and writers for each range
611
would compete for non-contiguous writes to multiple RocksDB logs.
612
613
In addition to the key/value pairs of the range itself, various range
614
metadata is maintained.
615
616
- range-spanning tree node links
617
618
- participating replicas
619
620
- consensus metadata
621
622
- split/merge activity
623
624
A really good reference on tuning Linux installations with RocksDB is
625
[here](http://docs.basho.com/riak/latest/ops/advanced/backends/leveldb/).
626
627
# Range Metadata
628
629
The default approximate size of a range is 64M (2\^26 B). In order to
630
support 1P (2\^50 B) of logical data, metadata is needed for roughly
631
2\^(50 - 26) = 2\^24 ranges. A reasonable upper bound on range metadata
633
locations and 220 bytes for the range key itself*). 2\^24 ranges \* 2\^8
634
B would require roughly 4G (2\^32 B) to store--too much to duplicate
635
between machines. Our conclusion is that range metadata must be
636
distributed for large installations.
637
638
To distribute the range metadata and keep key lookups relatively fast,
639
we use two levels of indirection. All of the range metadata sorts first
640
in our key-value map. We accomplish this by prefixing range metadata
641
with two null characters (*\0\0*). The *meta1* or *meta2* suffixes are
642
additionally appended to distinguish between the first level and second
644
we first locate the range information for the lower bound of
645
`\0\0meta1<key1>`, and then use that range to locate the lower bound
646
of `\0\0meta2<key1>`. The range specified there will indicate the
647
range location of `<key1>` (refer to examples below). Using two levels
648
of indirection, **our map can address approximately 2\^62 B of data, or
649
roughly 4E** (*each metadata range addresses 2\^(26-8) = 2\^18 ranges;
650
with two levels of indirection, we can address 2\^(18 + 18) = 2\^36
651
ranges; each range addresses 2\^26 B; total is 2\^(36+26) B = 2\^62 B =
652
4E*).
653
654
Note: we append the end key of each range to meta[12] records because
655
the RocksDB iterator only supports a Seek() interface which acts as a
656
Ceil(). Using the start key of the range would cause Seek() to find the
657
key *after* the meta indexing record we’re looking for, which would
658
result in having to back the iterator up, an option which is both less
659
efficient and not available in all cases.
660
661
The following example shows the directory structure for a map with
662
three ranges worth of data. Ellipses indicate additional key/value pairs to
663
fill an entire range of data. Except for the fact that splitting ranges
664
requires updates to the range metadata with knowledge of the metadata layout,
665
the range metadata itself requires no special treatment or bootstrapping.
666
667
**Range 0** (located on servers `dcrama1:8000`, `dcrama2:8000`,
668
`dcrama3:8000`)
669
670
- `\0\0meta1\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
671
- `\0\0meta2<lastkey0>`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
672
- `\0\0meta2<lastkey1>`: `dcrama4:8000`, `dcrama5:8000`, `dcrama6:8000`
673
- `\0\0meta2\xff`: `dcrama7:8000`, `dcrama8:8000`, `dcrama9:8000`
674
- ...
675
- `<lastkey0>`: `<lastvalue0>`
676
677
**Range 1** (located on servers `dcrama4:8000`, `dcrama5:8000`,
678
`dcrama6:8000`)
679
680
- ...
681
- `<lastkey1>`: `<lastvalue1>`
682
683
**Range 2** (located on servers `dcrama7:8000`, `dcrama8:8000`,
684
`dcrama9:8000`)
685
686
- ...
687
- `<lastkey2>`: `<lastvalue2>`
688
689
Consider a simpler example of a map containing less than a single
690
range of data. In this case, all range metadata and all data are
691
located in the same range:
692
693
**Range 0** (located on servers `dcrama1:8000`, `dcrama2:8000`,
694
`dcrama3:8000`)*
695
696
- `\0\0meta1\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
697
- `\0\0meta2\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
698
- `<key0>`: `<value0>`
699
- `...`
700
701
Finally, a map large enough to need both levels of indirection would
702
look like (note that instead of showing range replicas, this
703
example is simplified to just show range indexes):
704
705
**Range 0**
706
707
- `\0\0meta1<lastkeyN-1>`: Range 0
708
- `\0\0meta1\xff`: Range 1
709
- `\0\0meta2<lastkey1>`: Range 1
710
- `\0\0meta2<lastkey2>`: Range 2
711
- `\0\0meta2<lastkey3>`: Range 3
712
- ...
713
- `\0\0meta2<lastkeyN-1>`: Range 262143
714
715
**Range 1**
716
717
- `\0\0meta2<lastkeyN>`: Range 262144
718
- `\0\0meta2<lastkeyN+1>`: Range 262145
719
- ...
720
- `\0\0meta2\xff`: Range 500,000
721
- ...
722
- `<lastkey1>`: `<lastvalue1>`
723
724
**Range 2**
725
726
- ...
727
- `<lastkey2>`: `<lastvalue2>`
728
729
**Range 3**
730
731
- ...
732
- `<lastkey3>`: `<lastvalue3>`
733
734
**Range 262144**
735
736
- ...
737
- `<lastkeyN>`: `<lastvalueN>`
738
739
**Range 262145**
740
741
- ...
742
- `<lastkeyN+1>`: `<lastvalueN+1>`
743
744
Note that the choice of range `262144` is just an approximation. The
745
actual number of ranges addressable via a single metadata range is
746
dependent on the size of the keys. If efforts are made to keep key sizes
747
small, the total number of addressable ranges would increase and vice
748
versa.
749
750
From the examples above it’s clear that key location lookups require at
751
most three reads to get the value for `<key>`:
752
753
1. lower bound of `\0\0meta1<key>`
754
2. lower bound of `\0\0meta2<key>`,
755
3. `<key>`.
756
757
For small maps, the entire lookup is satisfied in a single RPC to Range 0. Maps
758
containing less than 16T of data would require two lookups. Clients cache both
759
levels of range metadata, and we expect that data locality for individual
760
clients will be high. Clients may end up with stale cache entries. If on a
761
lookup, the range consulted does not match the client’s expectations, the
762
client evicts the stale entries and possibly does a new lookup.
763
764
# Range-Spanning Binary Tree
765
766
A crucial enhancement to the organization of range metadata is to
767
augment the bi-level range metadata lookup with a minimum spanning tree,
768
implemented as a left-leaning red-black tree over all ranges in the map.
769
This tree structure allows the system to start at any key prefix and
770
efficiently traverse an arbitrary key range with minimal RPC traffic,
771
minimal fan-in and fan-out, and with bounded time complexity equal to
772
`2*log N` steps, where `N` is the total number of ranges in the system.
773
774
Unlike the range metadata rows prefixed with `\0\0meta[1|2]`, the
775
metadata for the range-spanning tree (e.g. parent range and left / right
776
child ranges) is stored directly at the ranges as non-map metadata. The
777
metadata for each node of the tree (e.g. links to parent range, left
778
child range, and right child range) is stored with the range metadata.
779
In effect, the tree metadata is stored implicitly. In order to traverse
780
the tree, for example, you’d need to query each range in turn for its
781
metadata.
782
783
Any time a range is split or merged, both the bi-level range lookup
784
metadata and the per-range binary tree metadata are updated as part of
785
the same distributed transaction. The total number of nodes involved in
786
the update is bounded by 2 + log N (i.e. 2 updates for meta1 and
787
meta2, and up to log N updates to balance the range-spanning tree).
788
The range corresponding to the root node of the tree is stored in
790
791
As an example, consider the following set of nine ranges and their
792
associated range-spanning tree:
793
794
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`.
795
796
797
798
The range-spanning tree has many beneficial uses in Cockroach. It makes
799
the problem of efficiently aggregating accounting information of
800
potentially vast ranges of data tractable. Imagine a subrange of data
801
over which accounting is being kept. For example, the *photos* table in
802
a public photo sharing site. To efficiently keep track of data about the
803
table (e.g. total size, number of rows, etc.), messages can be passed
804
first up the tree and then down to the left until updates arrive at the
805
key prefix under which accounting is aggregated. This makes worst case
806
number of hops for an update to propagate into the accounting totals
807
2 \* log N. A 64T database will require 1M ranges, meaning 40 hops
808
worst case. In our experience, accounting tasks over vast ranges of data
809
are most often map/reduce jobs scheduled with coarse-grained
810
periodicity. By contrast, we expect Cockroach to maintain statistics
811
with sub 10s accuracy and with minimal cycles and minimal IOPs.
812
813
Another use for the range-spanning tree is to push accounting, zones and
814
permissions configurations to all ranges. In the case of zones and
815
permissions, this is an efficient way to pass updated configuration
816
information with exponential fan-out. When adding accounting
817
configurations (i.e. specifying a new key prefix to track), the
818
implicated ranges are transactionally scanned and zero-state accounting
819
information is computed as well. Deleting accounting configurations is
820
similar, except accounting records are deleted.
821
822
Last but *not* least, the range-spanning tree provides a convenient
823
mechanism for planning and executing parallel queries. These provide the
824
basis for
825
[Dremel](http://static.googleusercontent.com/media/research.google.com/en/us/pubs/archive/36632.pdf)-like
826
query execution trees and it’s easy to imagine supporting a subset of
827
SQL or even javascript-based user functions for complex data analysis
828
tasks.
829
830
# Raft - Consistency of Range Replicas
831
832
Each range is configured to consist of three or more replicas. The
833
replicas in a range maintain their own instance of a distributed
834
consensus algorithm. We use the [*Raft consensus
835
algorithm*](https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf)
836
as it is simpler to reason about and includes a reference implementation
837
covering important details. Every write to replicas is logged twice.
838
Once to RocksDB’s internal log and once to levedb itself as part of the
839
Raft consensus log.
840
[ePaxos](https://www.cs.cmu.edu/~dga/papers/epaxos-sosp2013.pdf) has
841
promising performance characteristics for WAN-distributed replicas, but
842
it does not guarantee a consistent ordering between replicas.
843
844
Raft elects a relatively long-lived leader which must be involved to
845
propose writes. It heartbeats followers periodically to keep their logs
846
replicated. In the absence of heartbeats, followers become candidates
847
after randomized election timeouts and proceed to hold new leader
848
elections. Cockroach weights random timeouts such that the replicas with
849
shorter round trip times to peers are more likely to hold elections
850
first. Although only the leader can propose a new write, and as such
851
must be involved in any write to the consensus log, any replica can
852
service reads if the read is for a timestamp which the replica knows is
853
safe based on the last committed consensus write and the state of any
854
pending transactions.
855
856
Only the leader can propose a new write, but Cockroach accepts writes at
857
any replica. The replica merely forwards the write to the leader.
858
Instead of resending the write, the leader has only to acknowledge the
859
write to the forwarding replica using a log sequence number, as though
860
it were proposing it in the first place. The other replicas receive the
861
full write as though the leader were the originator.
862
863
Having a stable leader provides the choice of replica to handle
864
range-specific maintenance and processing tasks, such as delivering
865
pending message queues, handling splits and merges, rebalancing, etc.
866
867
# Splitting / Merging Ranges
868
869
RoachNodes split or merge ranges based on whether they exceed maximum or
870
minimum thresholds for capacity or load. Ranges exceeding maximums for
871
either capacity or load are split; ranges below minimums for *both*
872
capacity and load are merged.
873
874
Ranges maintain the same accounting statistics as accounting key
875
prefixes. These boil down to a time series of data points with minute
876
granularity. Everything from number of bytes to read/write queue sizes.
877
Arbitrary distillations of the accounting stats can be determined as the
878
basis for splitting / merging. Two sensical metrics for use with
879
split/merge are range size in bytes and IOps. A good metric for
880
rebalancing a replica from one node to another would be total read/write
881
queue wait times. These metrics are gossipped, with each range / node
882
passing along relevant metrics if they’re in the bottom or top of the
883
range it’s aware of.
884
885
A range finding itself exceeding either capacity or load threshold
886
splits. To this end, the range leader computes an appropriate split key
887
candidate and issues the split through Raft. In contrast to splitting,
888
merging requires a range to be below the minimum threshold for both
889
capacity *and* load. A range being merged chooses the smaller of the
890
ranges immediately preceding and succeeding it.
891
892
Splitting, merging, rebalancing and recovering all follow the same basic
893
algorithm for moving data between roach nodes. New target replicas are
894
created and added to the replica set of source range. Then each new
895
replica is brought up to date by either replaying the log in full or
896
copying a snapshot of the source replica data and then replaying the log
897
from the timestamp of the snapshot to catch up fully. Once the new
898
replicas are fully up to date, the range metadata is updated and old,
899
source replica(s) deleted if applicable.
900
901
**Coordinator** (leader replica)
902
906
only after being completed locally, are moved to new target replicas.
907
else if merging
908
Choose new replicas on same servers as target range replicas;
909
add to replica set.
910
else if rebalancing || recovering
911
Choose new replica(s) on least loaded servers; add to replica set.
912
```
913
914
**New Replica**
915
916
*Bring replica up to date:*
917
918
```
919
if all info can be read from replicated log
920
copy replicated log
921
else
922
snapshot source replica
923
send successive ReadRange requests to source replica
924
referencing snapshot
925
926
if merging
927
combine ranges on all replicas
928
else if rebalancing || recovering
929
remove old range replica(s)
930
```
931
932
RoachNodes split ranges when the total data in a range exceeds a
933
configurable maximum threshold. Similarly, ranges are merged when the
934
total data falls below a configurable minimum threshold.
935
936
**TBD: flesh this out**.
937
938
Ranges are rebalanced if a node determines its load or capacity is one
939
of the worst in the cluster based on gossipped load stats. A node with
940
spare capacity is chosen in the same datacenter and a special-case split
941
is done which simply duplicates the data 1:1 and resets the range
942
configuration metadata.
943
944
# Message Queues
945
946
Each range maintains an array of incoming message queues, referred to
947
here as **inboxes**. Additionally, each range maintains and *processes*
948
an array of outgoing message queues, referred to here as **outboxes**.
949
Both inboxes and outboxes are assigned to keys; messages can be sent or
950
received on behalf of any key. Inboxes and outboxes can contain any
951
number of pending messages.
952
953
Messages can be *deliverable*, or *executable.*
954
955
Deliverable messages are defined by Value objects - simple byte arrays -
956
that are delivered to a key’s inbox, awaiting collection by a client
957
invoking the ReapQueue operation. These are typically used by client
958
applications wishing to be notified of changes to an entry for further
959
processing, such as expensive offline operations like sending emails,
960
SMSs, etc.
961
962
Executable messages are *outgoing-only*, and are instances of
963
PutRequest,IncrementRequest, DeleteRequest, DeleteRangeRequest
965
executed when encountered. These are primarily useful when updates that
966
are nominally part of a transaction can tolerate asynchronous execution
967
(e.g. eventual consistency), and are otherwise too busy or numerous to
968
make including them in the original [distributed] transaction efficient.
969
Examples may include updates to the accounting for successive key
970
prefixes (potentially busy) or updates to a full-text index (potentially
971
numerous).
972
973
These two types of messages are enqueued in different outboxes too - see
974
key formats below.
975
976
At commit time, the range processing the transaction places messages
977
into a shared outbox located at the start of the range metadata. This is
978
effectively free as it’s part of the same consensus write for the range
979
as the COMMIT record. Outgoing messages are processed asynchronously by
980
the range. To make processing easy, all outboxes are co-located at the
981
start of the range. To make lookup easy, all inboxes are located
982
immediately after the recipient key. The leader replica of a range is
983
responsible for processing message queues.
984
985
A dispatcher polls a given range’s *deliverable message outbox*
986
periodically (configurable), and delivers those messages to the target
987
key’s inbox. The dispatcher is also woken up whenever a new message is
988
added to the outbox. A separate executor also polls the range’s
989
*executable message outbox* periodically as well (again, configurable),
991
new message is added to the outbox.
992
993
Formats follow in the table below. Notice that inbox messages for a
994
given key sort by the `<outbox-timestamp>`. This doesn’t provide a
995
precise ordering, but it does allow clients to scan messages in an
996
approximate ordering of when they were originally lodged with senders.
997
NTP offers walltime deltas to within 100s of milliseconds. The
998
`<sender-range-key>` suffix provides uniqueness.
999
1000
**Outbox**
1001
`<sender-range-key>deliverable-outbox:<recipient-key><outbox-timestamp>`
1002
`<sender-range-key>executable-outbox:<recipient-key><outbox-timestamp>`
1003
1004
**Inbox**
1005
`<recipient-key>inbox:<outbox-timestamp><sender-range-key>`
1006
1007
Messages are processed and then deleted as part of a single distributed
1008
transaction. The message will be executed or delivered exactly once,
1009
regardless of failures at either sender or receiver.
1010
1011
Delivered messages may be read by clients via the ReapQueue operation.
1012
This operation may only be used as part of a transaction. Clients should
1013
commit only after having processed the message. If the transaction is
1014
committed, scanned messages are automatically deleted. The operation
1015
name was chosen to reflect its mutating side effect. Deletion of read
1016
messages is mandatory because senders deliver messages asynchronously
1017
and a delay could cause insertion of messages at arbitrary points in the
1018
inbox queue. If clients require persistence, they should re-save read
1019
messages manually; the ReapQueue operation can be incorporated into
1020
normal transactional updates.
1021
1022
# Node Allocation (via Gossip)
1023
1024
New nodes must be allocated when a range is split. Instead of requiring
1025
every RoachNode to know about the status of all or even a large number
1026
of peer nodes --or-- alternatively requiring a specialized curator or
1027
master with sufficiently global knowledge, we use a gossip protocol to
1028
efficiently communicate only interesting information between all of the
1029
nodes in the cluster. What’s interesting information? One example would
1030
be whether a particular node has a lot of spare capacity. Each node,
1031
when gossiping, compares each topic of gossip to its own state. If its
1032
own state is somehow “more interesting” than the least interesting item
1033
in the topic it’s seen recently, it includes its own state as part of
1034
the next gossip session with a peer node. In this way, a node with
1035
capacity sufficiently in excess of the mean quickly becomes discovered
1036
by the entire cluster. To avoid piling onto outliers, nodes from the
1037
high capacity set are selected at random for allocation.
1038
1039
The gossip protocol itself contains two primary components:
1040
1041
- **Peer Selection**: each node maintains up to N peers with which it
1042
regularly communicates. It selects peers with an eye towards
1043
maximizing fanout. A peer node which itself communicates with an
1044
array of otherwise unknown nodes will be selected over one which
1045
communicates with a set containing significant overlap. Each time
1046
gossip is initiated, each nodes’ set of peers is exchanged. Each
1047
node is then free to incorporate the other’s peers as it sees fit.
1048
To avoid any node suffering from excess incoming requests, a node
1049
may refuse to answer a gossip exchange. Each node is biased
1050
towards answering requests from nodes without significant overlap
1051
and refusing requests otherwise.
1052
1053
Peers are efficiently selected using a heuristic as described in
1054
[Agarwal & Trachtenberg (2006)](https://drive.google.com/file/d/0B9GCVTp_FHJISmFRTThkOEZSM1U/edit?usp=sharing).
1055
1056
**TBD**: how to avoid partitions? Need to work out a simulation of
1057
the protocol to tune the behavior and see empirically how well it
1058
works.
1059
1060
- **Gossip Selection**: what to communicate. Gossip is divided into
1061
topics. Load characteristics (capacity per disk, cpu load, and
1062
state [e.g. draining, ok, failure]) are used to drive node
1063
allocation. Range statistics (range read/write load, missing
1064
replicas, unavailable ranges) and network topology (inter-rack
1065
bandwidth/latency, inter-datacenter bandwidth/latency, subnet
1066
outages) are used for determining when to split ranges, when to
1067
recover replicas vs. wait for network connectivity, and for
1068
debugging / sysops. In all cases, a set of minimums and a set of
1069
maximums is propagated; each node applies its own view of the
1070
world to augment the values. Each minimum and maximum value is
1071
tagged with the reporting node and other accompanying contextual
1072
information. Each topic of gossip has its own protobuf to hold the
1073
structured data. The number of items of gossip in each topic is
1074
limited by a configurable bound.
1075
1076
For efficiency, nodes assign each new item of gossip a sequence
1077
number and keep track of the highest sequence number each peer
1078
node has seen. Each round of gossip communicates only the delta
1079
containing new items.
1080
1081
# Node Accounting
1082
1083
The gossip protocol discussed in the previous section is useful to
1084
quickly communicate fragments of important information in a
1085
decentralized manner. However, complete accounting for each node is also
1086
stored to a central location, available to any dashboard process. This
1087
is done using the map itself. Each node periodically writes its state to
1088
the map with keys prefixed by `\0node`, similar to the first level of
1089
range metadata, but with an ‘`node`’ suffix. Each value is a protobuf
1090
containing the full complement of node statistics--everything
1091
communicated normally via the gossip protocol plus other useful, but
1092
non-critical data.
1093
1094
The range containing the first key in the node accounting table is
1095
responsible for gossiping the total count of nodes. This total count is
1096
used by the gossip network to most efficiently organize itself. In
1097
particular, the maximum number of hops for gossipped information to take
1098
before reaching a node is given by `ceil(log(node count) / log(max
1099
fanout)) + 1`.
1100
1101
# Key-prefix Accounting, Zones & Permissions
1102
1103
Arbitrarily fine-grained accounting and permissions are specified via
1104
key prefixes. Key prefixes can overlap, as is necessary for capturing
1105
hierarchical relationships. For illustrative purposes, let’s say keys
1106
specifying rows in a set of databases have the following format:
1107
1108
`<db>:<table>:<primary-key>[:<secondary-key>]`
1109
1110
In this case, we might collect accounting or specify permissions with
1111
key prefixes:
1112
1113
`db1`, `db1:user`, `db1:order`,
1114
1115
Accounting is kept for the entire map by default.
1116
1117
## Accounting
1118
to keep accounting for a range defined by a key prefix, an entry is created in
1119
the accounting system table. The format of accounting table keys is:
1120
1121
`\0acct<key-prefix>`
1122
1123
In practice, we assume each RoachNode capable of caching the
1124
entire accounting table as it is likely to be relatively small.
1125
1126
Accounting is kept for key prefix ranges with eventual consistency
1127
for efficiency. Updates to accounting values propagate through the
1128
system using the message queue facility if the accounting keys do
1129
not reside on the same range as ongoing activity (true for all but
1130
the smallest systems). There are two types of values which
1131
comprise accounting: counts and occurrences, for lack of better
1132
terms. Counts describe system state, such as the total number of
1133
bytes, rows, etc. Occurrences include transient performance and
1134
load metrics. Both types of accounting are captured as time series
1135
with minute granularity. The length of time accounting metrics are
1136
kept is configurable. Below are examples of each type of
1137
accounting value.
1138
1139
**System State Counters/Performance**
1140
1141
- Count of items (e.g. rows)
1142
- Total bytes
1143
- Total key bytes
1144
- Total value length
1145
- Queued message count
1146
- Queued message total bytes
1147
- Count of values \< 16B
1148
- Count of values \< 64B
1149
- Count of values \< 256B
1150
- Count of values \< 1K
1151
- Count of values \< 4K
1152
- Count of values \< 16K
1153
- Count of values \< 64K
1154
- Count of values \< 256K
1155
- Count of values \< 1M
1156
- Count of values \> 1M
1157
- Total bytes of accounting
1158
1159
1160
**Load Occurences**
1161
1162
Get op count
1163
Get total MB
1164
Put op count
1165
Put total MB
1166
Delete op count
1167
Delete total MB
1168
Delete range op count
1169
Delete range total MB
1170
Scan op count
1171
Scan op MB
1172
Split count
1173
Merge count
1174
1175
Because accounting information is kept as time series and over many
1176
possible metrics of interest, the data can become numerous. Accounting
1177
data are stored in the map near the key prefix described, in order to
1178
distribute load (for both aggregation and storage).
1179
1180
Accounting keys for system state have the form:
1181
`<key-prefix>|acctd<metric-name>*`. Notice the leading ‘pipe’
1182
character. It’s meant to sort the root level account AFTER any other
1183
system tables. They must increment the same underlying values as they
1184
are permanent counts, and not transient activity. Logic at the
1185
RoachNode takes care of snapshotting the value into an appropriately
1186
suffixed (e.g. with timestamp hour) multi-value time series entry.
1187
1188
Keys for perf/load metrics:
1189
`<key-prefix>acctd<metric-name><hourly-timestamp>`.
1190
1191
`<hourly-timestamp>`-suffixed accounting entries are multi-valued,
1192
containing a varint64 entry for each minute with activity during the
1193
specified hour.
1194
1195
To efficiently keep accounting over large key ranges, the task of
1196
aggregation must be distributed. If activity occurs within the same
1197
range as the key prefix for accounting, the updates are made as part
1198
of the consensus write. If the ranges differ, then a message is sent
1199
to the parent range to increment the accounting. If upon receiving the
1200
message, the parent range also does not include the key prefix, it in
1201
turn forwards it to its parent or left child in the balanced binary
1202
tree which is maintained to describe the range hierarchy. This limits
1203
the number of messages before an update is visible at the root to `2*log N`,
1204
where `N` is the number of ranges in the key prefix.
1205
1206
## Zones
1207
zones are stored in the map with keys prefixed by
1208
`\0zone` followed by the key prefix to which the zone
1209
configuration applies. Zone values specify a protobuf containing
1210
the datacenters from which replicas for ranges which fall under
1211
the zone must be chosen.
1212
1213
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`.
1214
1215
If zones are modified in situ, each RoachNode verifies the
1216
existing zones for its ranges against the zone configuration. If
1217
it discovers differences, it reconfigures ranges in the same way
1218
that it rebalances away from busy nodes, via special-case 1:1
1219
split to a duplicate range comprising the new configuration.
1220
1221
### Permissions
1222
permissions are stored in the map with keys prefixed by *\0perm* followed by
1223
the key prefix and user to which the specified permissions apply. The format of
1224
permissions keys is:
1225
1226
`\0perm<key-prefix><user>`
1227
1228
Permission values are a protobuf containing the permission details;
1229
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`.
1230
1231
A default system root permission is assumed for the entire map
1232
with an empty key prefix and read/write as true.
1233
1234
When determining whether or not to allow a read or a write a key
1235
value (e.g. `db1:user:1` for user `spencer`), a RoachNode would
1236
read the following permissions values:
1237
1238
```
1239
\0perm<db1:user:1>spencer
1240
\0perm<db1:user>spencer
1241
\0perm<db1>spencer
1242
\0perm<>spencer
1243
```
1244
1245
If any prefix in the hierarchy provides the required permission,
1246
the request is satisfied; otherwise, the request returns an
1247
error.
1248
1249
The priority for a user permission is used to order requests at
1250
Raft consensus ranges and for choosing an initial priority for
1251
distributed transactions. When scheduling operations at the Raft
1252
consensus range, all outstanding requests are ordered by key
1253
prefix and each assigned priorities according to key, user and
1254
arrival time. The next request is chosen probabilistically using
1255
priorities to weight the choice. Each key can have multiple
1256
priorities as they’re hierarchical (e.g. for /user/key, one
1257
priority for root ‘/’, and one for ‘/user/key’). The most general
1258
priority is used first. If two keys share the most general, then
1259
they’re compared with the next most general if applicable, and so on.
1260
1261
# Key-Value API
1262
1263
see the protobufs in [proto/](https://github.com/cockroachdb/cockroach/blob/master/proto),
1264
in particular [proto/api.proto](https://github.com/cockroachdb/cockroach/blob/master/proto/api.proto) and the comments within.
1265
1266
# Structured Data API
1267
1268
A preliminary design can be found in the [Go source documentation](http://godoc.org/github.com/cockroachdb/cockroach/structured).