Permalink
Apr 23, 2015
May 21, 2015
Apr 23, 2015
May 22, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
Apr 23, 2015
May 22, 2015
Apr 23, 2015
May 22, 2015
May 22, 2015
Apr 23, 2015
Apr 23, 2015
May 21, 2015
Apr 23, 2015
Apr 23, 2015
May 21, 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
Apr 23, 2015
Newer
100644
1262 lines (1047 sloc)
59.6 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
182
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”.
195
transaction.
196
197
2. Write an "intent" value for each datum being written as part of the
198
transaction. These are normal MVCC values, with the addition of a
199
special flag (i.e. “intent”) indicating that the value may be
201
the transaction id (unique and chosen at tx start time by client)
202
is stored with intent values. The tx id is used to refer to the
203
transaction table when there are conflicts and to make
204
tie-breaking decisions on ordering between identical timestamps.
207
selects the maximum from amongst all writes as the final commit
208
timestamp.
209
210
Each range maintains a small (i.e. latest 10s of read timestamps),
211
*in-memory* cache from key to the latest timestamp at which the
212
key(s) were read. This *latest-read-cache* is consulted on each
213
write. If the write’s candidate timestamp is earlier than the low
214
water mark on the cache itself (i.e. its last evicted timestamp)
215
or if the key being written has a read timestamp later than the
216
write’s candidate timestamp, this later timestamp value is
217
returned with the write. The cache’s entries are evicted oldest
218
timestamp first, updating low water mark as appropriate. If a new
219
range replica leader is elected, it sets the low water mark for
220
the cache to the current wall time + ε (ε = 99^th^ percentile
221
clock skew).
222
223
3. Commit the transaction by updating its entry in the system
224
transaction table (keys prefixed by *\0tx*). The value of the
225
commit entry contains the candidate timestamp (increased as
226
necessary to accommodate any latest read timestamps). Note that
227
the transaction is considered fully committed at this point and
228
control may be returned to the client.
229
230
In the case of an SI transaction, a commit timestamp which was
231
increased to accommodate concurrent readers is perfectly
232
acceptable and the commit may continue. For SSI transactions,
233
however, a gap between candidate and commit timestamps
234
necessitates transaction restart (note: restart is different than
235
abort--see below).
236
237
After the transaction is committed, all written intents are upgraded
238
in parallel by removing the “intent” flag. The transaction is
239
considered fully committed before this step and does not wait for
240
it to return control to the transaction coordinator.
241
242
In the absence of conflicts, this is the end. Nothing else is necessary
243
to ensure the correctness of the system.
244
245
**Conflict Resolution**
246
247
Things get more interesting when a reader or writer encounters an intent
248
record or newly-committed value in a location that it needs to read or
249
write. This is a conflict, usually causing either of the transactions to
250
abort or restart depending on the type of conflict.
251
252
***Transaction restart:***
253
254
This is the usual (and more efficient) type of behaviour and is used
255
except when the transaction was aborted (for instance by another
256
transaction).
257
In effect, that reduces to two cases, the first being the one outlined
258
above: An SSI transaction that finds (upon attempting to commit) that
259
its commit timestamp has been pushed. In the second case, a transaction
260
actively encounters a conflict, that is, one of its readers or writers
261
runs encounters data that necessitate conflict resolution.
262
263
When a transaction restarts, it changes its priority and/or moves its
264
timestamp forward depending on data tied to the conflict, and
265
begins anew reusing the same tx id. Since the set of keys
266
being written changes between restarts, a set of keys written during
268
bogus intents. As the client replays the transaction from the
269
beginning, it removes keys from the set as it writes them again. The remaining
270
keys in the set--should the transaction run to completion--refer bogus
271
write intents which must be deleted *before* the transaction commits
272
Many transactions will end up with an empty set implying no bogus write
273
intents.
274
275
***Transaction abort:***
276
277
This is the case in which a transaction, upon reading its transaction
278
table entry, finds that it has been aborted. In this case, the
279
transaction can not reuse its intents; it returns control to the client
280
before cleaning them up (other readers and writers would clean up
281
dangling intents as they encounter them) but will make an effort to
282
clean up after itself. The next attempt (if applicable) then runs as a
286
287
There are several scenarios in which transactions interact:
288
289
- **Reader encounters write intent or value with newer timestamp far
290
enough in the future**: This is not a conflict. The reader is free
291
to proceed; after all, it will be reading an older version of the
292
value and so does not conflict. Recall that the write intent may
293
be committed with a later timestamp than its candidate; it will
294
never commit with an earlier one. **Side note**: if a SI transaction
295
reader finds an intent with a newer timestamp which the reader’s own
296
transaction has written, the reader always returns that value.
297
298
- **Reader encounters write intent or value with newer timestamp in the
299
near future:** In this case, we have to be careful. The newer
300
intent may, in absolute terms, have happened in our read's past if
301
the clock of the writer is ahead of the node serving the values.
302
In that case, we would need to take this value into account, but
303
we just don't know. Hence the transaction restarts, using instead
304
a future timestamp (but remembering a maximum timestamp used to
305
limit the uncertainty window to the maximum clock skew). In fact,
306
this is optimized further; see the details under "choosing a time
307
stamp" below.
308
309
- **Reader encounters write intent with older timestamp**: the reader
310
must follow the intent’s transaction id to the transaction table.
311
If the transaction has already been committed, then the reader can
312
just read the value. If the write transaction has not yet been
313
committed, then the reader has two options. If the write conflict
314
is from an SI transaction, the reader can *push that transaction's
315
commit timestamp into the future* (and consequently not have to
316
read it). This is simple to do: the reader just updates the
317
transaction’s commit timestamp to indicate that when/if the
318
transaction does commit, it should use a timestamp *at least* as
319
high. However, if the write conflict is from an SSI transaction,
320
the reader must compare priorities. If it has the higher priority,
321
it pushes the transaction’s commit timestamp, as with SI (that
322
transaction will then notice its timestamp has been pushed, and
323
restart). If it has the lower priority, it retries itself using as
324
a new priority `max(new random priority, conflicting txn’s
326
327
- **Writer encounters uncommitted write intent**:
328
If the write intent has been written by a transaction with a lower
329
priority writer aborts the conflicting transaction. if the intent has
330
a higher or equal priority the transaction retries, using as a new
331
priority *max(new random priority, conflicting txn’s priority - 1)*;
332
the retry occurs after a short, randomized backoff interval.
333
335
The transaction restarts. On restart, the same priority is reused,
336
but the candidate timestamp is moved forward to the encountered
337
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 cleanup write intents upon transaction
346
completion.
347
348
If a transaction is completed successfully, all intents are upgraded to
349
committed. In the event a transaction is aborted, all written intents
350
are deleted. The client proxy doesn’t guarantee it will cleanup intents;
351
but dangling intents are upgraded or deleted when encountered by future
352
readers and writers and the system does not depend on their timely
353
cleanup for correctness.
354
355
In the event the client proxy restarts before the pending transaction is
356
completed, the dangling transaction would continue to live in the
357
transaction table until aborted by another transaction. Transactions
358
heartbeat the transaction table every five seconds by default.
359
Transactions encountered by readers or writers with dangling intents
360
which haven’t been heartbeat within the required interval are aborted.
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 the same.
428
Time spent retrying because of reading recently committed data has an
429
upper bound of `ε`. In fact, this is further optimized: upon restarting,
430
the transaction not only takes into account the timestamp of the future
431
value, but the timestamp of the node at the time of the uncertain read.
432
The larger of those two timestamps (typically, the latter) is used to
433
bump up the read timestamp, and additionally the node is marked as
434
“certain”. This means that for future reads to that node within the
435
transaction, we can set `MaxTimestamp = Read Timestamp` (and hence avoid
436
further uncertainty restarts). Correctness follows from the fact that we
437
know that at the time of the read, there exists no version of any key on
438
that node with a higher timestamp; if we ran into one during a future
439
read, that node would have happened (in absolute time) after our
440
transaction started.
441
442
We expect retries will be rare, but this assumption may need to be
443
revisited if retries become problematic. Note that this problem does not
444
apply to historical reads. An alternate approach which does not require
445
retries would be to make a round to all node participants in advance and
446
choose the highest reported node wall time as the timestamp. However,
447
knowing which nodes will be accessed in advance is difficult and
448
potentially limiting. Cockroach could also potentially use a global
449
clock (Google did this with Pinax TODO: link to paper), which would be
450
feasible for smaller, geographically-proximate clusters.
451
452
# Linearizability
453
454
First a word about [***Spanner***](http://research.google.com/archive/spanner.html).
455
By combining judicious use of wait intervals with accurate time signals,
456
Spanner provides a global ordering between any two non-overlapping transactions
457
(in absolute time) with \~14ms latencies. Put another way:
458
Spanner guarantees that if a transaction T<sub>1</sub> commits (in absolute time)
459
before another transaction T<sub>2</sub> starts, then T<sub>1</sub>'s assigned commit
460
timestamp is smaller than T<sub>2</sub>'s. Using atomic clocks and GPS receivers,
461
Spanner reduces their clock skew uncertainty to \< 10ms (`ε`). To make
462
good on the promised guarantee, transactions must take at least double
463
the clock skew uncertainty interval to commit (`2ε`). See [*this
464
article*](http://www.cs.cornell.edu/~ie53/publications/DC-col51-Sep13.pdf)
465
for a helpful overview of Spanner’s concurrency control.
466
467
Cockroach could make the same guarantees without specialized hardware,
468
at the expense of longer wait times. If servers in the cluster were
469
configured to work only with NTP, transaction wait times would likely to
470
be in excess of 150ms. For wide-area zones, this would be somewhat
471
mitigated by overlap from cross datacenter link latencies. If clocks
472
were made more accurate, the minimal limit for commit latencies would
473
improve.
474
475
However, let’s take a step back and evaluate whether Spanner’s external
476
consistency guarantee is worth the automatic commit wait. First, if the
477
commit wait is omitted completely, the system still yields a consistent
478
view of the map at an arbitrary timestamp. However with clock skew, it
479
would become possible for commit timestamps on non-overlapping but
480
causally related transactions to suffer temporal reverse. In other
481
words, the following scenario is possible for a client without global
482
ordering:
483
484
- Start transaction T<sub>1</sub> to modify value `x` with commit time *s<sub>1</sub>*
485
486
- On commit of T<sub>1</sub>, start T<sub>2</sub> to modify value `y` with commit time
487
\> s<sub>2</sub>
488
489
- Read `x` and `y` and discover that s<sub>1</sub> \> s<sub>2</sub> (**!**)
490
491
The external consistency which Spanner guarantees is referred to as
492
**linearizability**. It goes beyond serializability by preserving
493
information about the causality inherent in how external processes
494
interacted with the database. The strength of Spanner’s guarantee can be
495
formulated as follows: any two processes, with clock skew within
496
expected bounds, may independently record their wall times for the
497
completion of transaction T<sub>1</sub> (T<sub>1</sub><sup>end</sup>) and start of transaction
498
T<sub>2</sub> (T<sub>2</sub><sup>start</sup>) respectively, and if later
499
compared such that T<sub>1</sub><sup>end</sup> \< T<sub>2</sub><sup>start</sup>,
500
then commit timestamps s<sub>1</sub> \< s<sub>2</sub>.
501
This guarantee is broad enough to completely cover all cases of explicit
502
causality, in addition to covering any and all imaginable scenarios of implicit
503
causality.
504
505
Our contention is that causality is chiefly important from the
506
perspective of a single client or a chain of successive clients (*if a
507
tree falls in the forest and nobody hears…*). As such, Cockroach
508
provides two mechanisms to provide linearizability for the vast majority
509
of use cases without a mandatory transaction commit wait or an elaborate
510
system to minimize clock skew.
511
512
1. Clients provide the highest transaction commit timestamp with
513
> successive transactions. This allows node clocks from previous
514
> transactions to effectively participate in the formulation of the
515
> commit timestamp for the current transaction. This guarantees
516
> linearizability for transactions committed by this client.
517
>
518
> Newly launched clients wait at least 2 \* ε from process start
519
> time before beginning their first transaction. This preserves the
520
> same property even on client restart, and the wait will be
521
> mitigated by process initialization.
522
>
523
> All causally-related events within Cockroach maintain
524
> linearizability. Message queues, for example, guarantee that the
525
> receipt timestamp is greater than send timestamp, and that
526
> delivered messages may not be reaped until after the commit wait.
527
528
2. Committed transactions respond with a commit wait parameter which
529
> represents the remaining time in the nominal commit wait. This
530
> will typically be less than the full commit wait as the consensus
531
> write at the coordinator accounts for a portion of it.
532
>
533
> Clients taking any action outside of another Cockroach transaction
534
> (e.g. writing to another distributed system component) can either
535
> choose to wait the remaining interval before proceeding, or
536
> alternatively, pass the wait and/or commit timestamp to the
537
> execution of the outside action for its consideration. This pushes
538
> the burden of linearizability to clients, but is a useful tool in
539
> mitigating commit latencies if the clock skew is potentially
540
> large. This functionality can be used for ordering in the face of
541
> backchannel dependencies as mentioned in the
542
> [AugmentedTime](http://www.cse.buffalo.edu/~demirbas/publications/augmentedTime.pdf)
543
> paper.
544
545
Using these mechanisms in place of commit wait, Cockroach’s guarantee can be
546
formulated as follows: any process which signals the start of transaction
547
T<sub>2</sub> (T<sub>2</sub><sup>start</sup>) after the completion of
548
transaction T<sub>1</sub> (T<sub>1</sub><sup>end</sup>), will have commit
549
timestamps such thats<sub>1</sub> \< s<sub>2</sub>.
550
551
# Logical Map Content
552
553
Logically, the map contains a series of reserved system key / value
554
pairs covering accounting, range metadata, node accounting and
555
permissions before the actual key / value pairs for non-system data
556
(e.g. the actual meat of the map).
557
558
- `\0\0meta1` Range metadata for location of `\0\0meta2`.
559
- `\0\0meta1<key1>` Range metadata for location of `\0\0meta2<key1>`.
560
- ...
561
- `\0\0meta1<keyN>`: Range metadata for location of `\0\0meta2<keyN>`.
562
- `\0\0meta2`: Range metadata for location of first non-range metadata key.
563
- `\0\0meta2<key1>`: Range metadata for location of `<key1>`.
564
- ...
565
- `\0\0meta2<keyN>`: Range metadata for location of `<keyN>`.
566
- `\0acct<key0>`: Accounting for key prefix key0.
567
- ...
568
- `\0acct<keyN>`: Accounting for key prefix keyN.
569
- `\0node<node-address0>`: Accounting data for node 0.
570
- ...
571
- `\0node<node-addressN>`: Accounting data for node N.
572
- `\0perm<key0><user0>`: Permissions for user0 for key prefix key0.
573
- ...
574
- `\0perm<keyN><userN>`: Permissions for userN for key prefix keyN.
575
- `\0tree_root`: Range key for root of range-spanning tree.
576
- `\0tx<tx-id0>`: Transaction record for transaction 0.
577
- ...
578
- `\0tx<tx-idN>`: Transaction record for transaction N.
579
- `\0zone<key0>`: Zone information for key prefix key0.
580
- ...
581
- `\0zone<keyN>`: Zone information for key prefix keyN.
582
- `<>acctd<metric0>`: Accounting data for Metric 0 for empty key prefix.
583
- ...
584
- `<>acctd<metricN>`: Accounting data for Metric N for empty key prefix.
585
- `<key0>`: `<value0>` The first user data key.**
586
- ...
587
- `<keyN>`: `<valueN>` The last user data key.**
588
589
There are some additional system entries sprinkled amongst the
590
non-system keys. See the Key-Prefix Accounting section in this document
591
for further details.
592
593
# Node Storage
594
595
Nodes maintain a separate instance of RocksDB for each disk. Each
596
RocksDB instance hosts any number of ranges. RPCs arriving at a
597
RoachNode are multiplexed based on the disk name to the appropriate
598
RocksDB instance. A single instance per disk is used to avoid
599
contention. If every range maintained its own RocksDB, global management
600
of available cache memory would be impossible and writers for each range
601
would compete for non-contiguous writes to multiple RocksDB logs.
602
603
In addition to the key/value pairs of the range itself, various range
604
metadata is maintained.
605
606
- range-spanning tree node links
607
608
- participating replicas
609
610
- consensus metadata
611
612
- split/merge activity
613
614
A really good reference on tuning Linux installations with RocksDB is
615
[here](http://docs.basho.com/riak/latest/ops/advanced/backends/leveldb/).
616
617
# Range Metadata
618
619
The default approximate size of a range is 64M (2\^26 B). In order to
620
support 1P (2\^50 B) of logical data, metadata is needed for roughly
621
2\^(50 - 26) = 2\^24 ranges. A reasonable upper bound on range metadata
622
size is roughly 256 bytes (*3 \* 12 bytes for the triplicated node
623
locations and 220 bytes for the range key itself*). 2\^24 ranges \* 2\^8
624
B would require roughly 4G (2\^32 B) to store--too much to duplicate
625
between machines. Our conclusion is that range metadata must be
626
distributed for large installations.
627
628
To distribute the range metadata and keep key lookups relatively fast,
629
we use two levels of indirection. All of the range metadata sorts first
630
in our key-value map. We accomplish this by prefixing range metadata
631
with two null characters (*\0\0*). The *meta1* or *meta2* suffixes are
632
additionally appended to distinguish between the first level and second
633
level of **`r`**a***ng***e metadata. In order to do a lookup for *key1*,
634
we first locate the range information for the lower bound of
635
`\0\0meta1<key1>`, and then use that range to locate the lower bound
636
of `\0\0meta2<key1>`. The range specified there will indicate the
637
range location of `<key1>` (refer to examples below). Using two levels
638
of indirection, **our map can address approximately 2\^62 B of data, or
639
roughly 4E** (*each metadata range addresses 2\^(26-8) = 2\^18 ranges;
640
with two levels of indirection, we can address 2\^(18 + 18) = 2\^36
641
ranges; each range addresses 2\^26 B; total is 2\^(36+26) B = 2\^62 B =
642
4E*).
643
644
Note: we append the end key of each range to meta[12] records because
645
the RocksDB iterator only supports a Seek() interface which acts as a
646
Ceil(). Using the start key of the range would cause Seek() to find the
647
key *after* the meta indexing record we’re looking for, which would
648
result in having to back the iterator up, an option which is both less
649
efficient and not available in all cases.
650
651
The following example shows the directory structure for a map with
652
three ranges worth of data. The key/values in red show range
653
metadata. The key/values in black show actual data. Ellipses
654
indicate additional key/value pairs to fill out entire range of
655
data. Except for the fact that splitting ranges requires updates
656
to the range metadata with knowledge of the metadata layout, the
657
range metadata itself requires no special treatment or
658
bootstrapping.
659
660
**Range 0** (located on servers `dcrama1:8000`, `dcrama2:8000`,
661
`dcrama3:8000`)
662
663
- `\0\0meta1\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
664
- `\0\0meta2<lastkey0>`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
665
- `\0\0meta2<lastkey1>`: `dcrama4:8000`, `dcrama5:8000`, `dcrama6:8000`
666
- `\0\0meta2\xff`: `dcrama7:8000`, `dcrama8:8000`, `dcrama9:8000`
667
- ...
668
- `<lastkey0>`: `<lastvalue0>`
669
670
**Range 1** (located on servers `dcrama4:8000`, `dcrama5:8000`,
671
`dcrama6:8000`)
672
673
- ...
674
- `<lastkey1>`: `<lastvalue1>`
675
676
**Range 2** (located on servers `dcrama7:8000`, `dcrama8:8000`,
677
`dcrama9:8000`)
678
679
- ...
680
- `<lastkey2>`: `<lastvalue2>`
681
682
Consider a simpler example of a map containing less than a single
683
range of data. In this case, all range metadata and all data are
684
located in the same range:
685
686
**Range 0** (located on servers `dcrama1:8000`, `dcrama2:8000`,
687
`dcrama3:8000`)*
688
689
- `\0\0meta1\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
690
- `\0\0meta2\xff`: `dcrama1:8000`, `dcrama2:8000`, `dcrama3:8000`
691
- `<key0>`: `<value0>`
692
- `...`
693
694
Finally, a map large enough to need both levels of indirection would
695
look like (note that instead of showing range replicas, this
696
example is simplified to just show range indexes):
697
698
**Range 0**
699
700
- `\0\0meta1<lastkeyN-1>`: Range 0
701
- `\0\0meta1\xff`: Range 1
702
- `\0\0meta2<lastkey1>`: Range 1
703
- `\0\0meta2<lastkey2>`: Range 2
704
- `\0\0meta2<lastkey3>`: Range 3
705
- ...
706
- `\0\0meta2<lastkeyN-1>`: Range 262143
707
708
**Range 1**
709
710
- `\0\0meta2<lastkeyN>`: Range 262144
711
- `\0\0meta2<lastkeyN+1>`: Range 262145
712
- ...
713
- `\0\0meta2\xff`: Range 500,000
714
- ...
715
- `<lastkey1>`: `<lastvalue1>`
716
717
**Range 2**
718
719
- ...
720
- `<lastkey2>`: `<lastvalue2>`
721
722
**Range 3**
723
724
- ...
725
- `<lastkey3>`: `<lastvalue3>`
726
727
**Range 262144**
728
729
- ...
730
- `<lastkeyN>`: `<lastvalueN>`
731
732
**Range 262145**
733
734
- ...
735
- `<lastkeyN+1>`: `<lastvalueN+1>`
736
737
Note that the choice of range `262144` is just an approximation. The
738
actual number of ranges addressable via a single metadata range is
739
dependent on the size of the keys. If efforts are made to keep key sizes
740
small, the total number of addressable ranges would increase and vice
741
versa.
742
743
From the examples above it’s clear that key location lookups require at
744
most three reads to get the value for `<key>`:
745
746
1. lower bound of `\0\0meta1<key>`
747
2. lower bound of `\0\0meta2<key>`,
748
3. `<key>`.
749
750
For small maps, the entire lookup is satisfied in a single RPC to Range 0. Maps
751
containing less than 16T of data would require two lookups. Clients cache both
752
levels of range metadata, and we expect that data locality for individual
753
clients will be high. Clients may end up with stale cache entries. If on a
754
lookup, the range consulted does not match the client’s expectations, the
755
client evicts the stale entries and possibly does a new lookup.
756
757
# Range-Spanning Binary Tree
758
759
A crucial enhancement to the organization of range metadata is to
760
augment the bi-level range metadata lookup with a minimum spanning tree,
761
implemented as a left-leaning red-black tree over all ranges in the map.
762
This tree structure allows the system to start at any key prefix and
763
efficiently traverse an arbitrary key range with minimal RPC traffic,
764
minimal fan-in and fan-out, and with bounded time complexity equal to
765
`2*log N` steps, where `N` is the total number of ranges in the system.
766
767
Unlike the range metadata rows prefixed with `\0\0meta[1|2]`, the
768
metadata for the range-spanning tree (e.g. parent range and left / right
769
child ranges) is stored directly at the ranges as non-map metadata. The
770
metadata for each node of the tree (e.g. links to parent range, left
771
child range, and right child range) is stored with the range metadata.
772
In effect, the tree metadata is stored implicitly. In order to traverse
773
the tree, for example, you’d need to query each range in turn for its
774
metadata.
775
776
Any time a range is split or merged, both the bi-level range lookup
777
metadata and the per-range binary tree metadata are updated as part of
778
the same distributed transaction. The total number of nodes involved in
779
the update is bounded by 2 + log N (i.e. 2 updates for meta1 and
780
meta2, and up to log N updates to balance the range-spanning tree).
781
The range corresponding to the root node of the tree is stored in
783
784
As an example, consider the following set of nine ranges and their
785
associated range-spanning tree:
786
787
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`.
788
789
790
791
The range-spanning tree has many beneficial uses in Cockroach. It makes
792
the problem of efficiently aggregating accounting information of
793
potentially vast ranges of data tractable. Imagine a subrange of data
794
over which accounting is being kept. For example, the *photos* table in
795
a public photo sharing site. To efficiently keep track of data about the
796
table (e.g. total size, number of rows, etc.), messages can be passed
797
first up the tree and then down to the left until updates arrive at the
798
key prefix under which accounting is aggregated. This makes worst case
799
number of hops for an update to propagate into the accounting totals
800
2 \* log N. A 64T database will require 1M ranges, meaning 40 hops
801
worst case. In our experience, accounting tasks over vast ranges of data
802
are most often map/reduce jobs scheduled with coarse-grained
803
periodicity. By contrast, we expect Cockroach to maintain statistics
804
with sub 10s accuracy and with minimal cycles and minimal IOPs.
805
806
Another use for the range-spanning tree is to push accounting, zones and
807
permissions configurations to all ranges. In the case of zones and
808
permissions, this is an efficient way to pass updated configuration
809
information with exponential fan-out. When adding accounting
810
configurations (i.e. specifying a new key prefix to track), the
811
implicated ranges are transactionally scanned and zero-state accounting
812
information is computed as well. Deleting accounting configurations is
813
similar, except accounting records are deleted.
814
815
Last but *not* least, the range-spanning tree provides a convenient
816
mechanism for planning and executing parallel queries. These provide the
817
basis for
818
[Dremel](http://static.googleusercontent.com/media/research.google.com/en/us/pubs/archive/36632.pdf)-like
819
query execution trees and it’s easy to imagine supporting a subset of
820
SQL or even javascript-based user functions for complex data analysis
821
tasks.
822
823
# Raft - Consistency of Range Replicas
824
825
Each range is configured to consist of three or more replicas. The
826
replicas in a range maintain their own instance of a distributed
827
consensus algorithm. We use the [*Raft consensus
828
algorithm*](https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf)
829
as it is simpler to reason about and includes a reference implementation
830
covering important details. Every write to replicas is logged twice.
831
Once to RocksDB’s internal log and once to levedb itself as part of the
832
Raft consensus log.
833
[ePaxos](https://www.cs.cmu.edu/~dga/papers/epaxos-sosp2013.pdf) has
834
promising performance characteristics for WAN-distributed replicas, but
835
it does not guarantee a consistent ordering between replicas.
836
837
Raft elects a relatively long-lived leader which must be involved to
838
propose writes. It heartbeats followers periodically to keep their logs
839
replicated. In the absence of heartbeats, followers become candidates
840
after randomized election timeouts and proceed to hold new leader
841
elections. Cockroach weights random timeouts such that the replicas with
842
shorter round trip times to peers are more likely to hold elections
843
first. Although only the leader can propose a new write, and as such
844
must be involved in any write to the consensus log, any replica can
845
service reads if the read is for a timestamp which the replica knows is
846
safe based on the last committed consensus write and the state of any
847
pending transactions.
848
849
Only the leader can propose a new write, but Cockroach accepts writes at
850
any replica. The replica merely forwards the write to the leader.
851
Instead of resending the write, the leader has only to acknowledge the
852
write to the forwarding replica using a log sequence number, as though
853
it were proposing it in the first place. The other replicas receive the
854
full write as though the leader were the originator.
855
856
Having a stable leader provides the choice of replica to handle
857
range-specific maintenance and processing tasks, such as delivering
858
pending message queues, handling splits and merges, rebalancing, etc.
859
860
# Splitting / Merging Ranges
861
862
RoachNodes split or merge ranges based on whether they exceed maximum or
863
minimum thresholds for capacity or load. Ranges exceeding maximums for
864
either capacity or load are split; ranges below minimums for *both*
865
capacity and load are merged.
866
867
Ranges maintain the same accounting statistics as accounting key
868
prefixes. These boil down to a time series of data points with minute
869
granularity. Everything from number of bytes to read/write queue sizes.
870
Arbitrary distillations of the accounting stats can be determined as the
871
basis for splitting / merging. Two sensical metrics for use with
872
split/merge are range size in bytes and IOps. A good metric for
873
rebalancing a replica from one node to another would be total read/write
874
queue wait times. These metrics are gossipped, with each range / node
875
passing along relevant metrics if they’re in the bottom or top of the
876
range it’s aware of.
877
878
A range finding itself exceeding either capacity or load threshold
879
splits. To this end, the range leader computes an appropriate split key
880
candidate and issues the split through Raft. In contrast to splitting,
881
merging requires a range to be below the minimum threshold for both
882
capacity *and* load. A range being merged chooses the smaller of the
883
ranges immediately preceding and succeeding it.
884
885
Splitting, merging, rebalancing and recovering all follow the same basic
886
algorithm for moving data between roach nodes. New target replicas are
887
created and added to the replica set of source range. Then each new
888
replica is brought up to date by either replaying the log in full or
889
copying a snapshot of the source replica data and then replaying the log
890
from the timestamp of the snapshot to catch up fully. Once the new
891
replicas are fully up to date, the range metadata is updated and old,
892
source replica(s) deleted if applicable.
893
894
**Coordinator** (leader replica)
895
899
only after being completed locally, are moved to new target replicas.
900
else if merging
901
Choose new replicas on same servers as target range replicas;
902
add to replica set.
903
else if rebalancing || recovering
904
Choose new replica(s) on least loaded servers; add to replica set.
905
```
906
907
**New Replica**
908
909
*Bring replica up to date:*
910
911
```
912
if all info can be read from replicated log
913
copy replicated log
914
else
915
snapshot source replica
916
send successive ReadRange requests to source replica
917
referencing snapshot
918
919
if merging
920
combine ranges on all replicas
921
else if rebalancing || recovering
922
remove old range replica(s)
923
```
924
925
RoachNodes split ranges when the total data in a range exceeds a
926
configurable maximum threshold. Similarly, ranges are merged when the
927
total data falls below a configurable minimum threshold.
928
929
**TBD: flesh this out**.
930
931
Ranges are rebalanced if a node determines its load or capacity is one
932
of the worst in the cluster based on gossipped load stats. A node with
933
spare capacity is chosen in the same datacenter and a special-case split
934
is done which simply duplicates the data 1:1 and resets the range
935
configuration metadata.
936
937
# Message Queues
938
939
Each range maintains an array of incoming message queues, referred to
940
here as **inboxes**. Additionally, each range maintains and *processes*
941
an array of outgoing message queues, referred to here as **outboxes**.
942
Both inboxes and outboxes are assigned to keys; messages can be sent or
943
received on behalf of any key. Inboxes and outboxes can contain any
944
number of pending messages.
945
946
Messages can be *deliverable*, or *executable.*
947
948
Deliverable messages are defined by Value objects - simple byte arrays -
949
that are delivered to a key’s inbox, awaiting collection by a client
950
invoking the ReapQueue operation. These are typically used by client
951
applications wishing to be notified of changes to an entry for further
952
processing, such as expensive offline operations like sending emails,
953
SMSs, etc.
954
955
Executable messages are *outgoing-only*, and are instances of
956
PutRequest,IncrementRequest, DeleteRequest, DeleteRangeRequest
957
orAccountingRequest. Rather than being delivered to a key’s inbox, are
958
executed when encountered. These are primarily useful when updates that
959
are nominally part of a transaction can tolerate asynchronous execution
960
(e.g. eventual consistency), and are otherwise too busy or numerous to
961
make including them in the original [distributed] transaction efficient.
962
Examples may include updates to the accounting for successive key
963
prefixes (potentially busy) or updates to a full-text index (potentially
964
numerous).
965
966
These two types of messages are enqueued in different outboxes too - see
967
key formats below.
968
969
At commit time, the range processing the transaction places messages
970
into a shared outbox located at the start of the range metadata. This is
971
effectively free as it’s part of the same consensus write for the range
972
as the COMMIT record. Outgoing messages are processed asynchronously by
973
the range. To make processing easy, all outboxes are co-located at the
974
start of the range. To make lookup easy, all inboxes are located
975
immediately after the recipient key. The leader replica of a range is
976
responsible for processing message queues.
977
978
A dispatcher polls a given range’s *deliverable message outbox*
979
periodically (configurable), and delivers those messages to the target
980
key’s inbox. The dispatcher is also woken up whenever a new message is
981
added to the outbox. A separate executor also polls the range’s
982
*executable message outbox* periodically as well (again, configurable),
983
and executes those commands. The exeecutor, too, is woken up whenever a
984
new message is added to the outbox.
985
986
Formats follow in the table below. Notice that inbox messages for a
987
given key sort by the `<outbox-timestamp>`. This doesn’t provide a
988
precise ordering, but it does allow clients to scan messages in an
989
approximate ordering of when they were originally lodged with senders.
990
NTP offers walltime deltas to within 100s of milliseconds. The
991
`<sender-range-key>` suffix provides uniqueness.
992
993
**Outbox**
994
`<sender-range-key>deliverable-outbox:<recipient-key><outbox-timestamp>`
995
`<sender-range-key>executable-outbox:<recipient-key><outbox-timestamp>`
996
997
**Inbox**
998
`<recipient-key>inbox:<outbox-timestamp><sender-range-key>`
999
1000
Messages are processed and then deleted as part of a single distributed
1001
transaction. The message will be executed or delivered exactly once,
1002
regardless of failures at either sender or receiver.
1003
1004
Delivered messages may be read by clients via the ReapQueue operation.
1005
This operation may only be used as part of a transaction. Clients should
1006
commit only after having processed the message. If the transaction is
1007
committed, scanned messages are automatically deleted. The operation
1008
name was chosen to reflect its mutating side effect. Deletion of read
1009
messages is mandatory because senders deliver messages asynchronously
1010
and a delay could cause insertion of messages at arbitrary points in the
1011
inbox queue. If clients require persistence, they should re-save read
1012
messages manually; the ReapQueue operation can be incorporated into
1013
normal transactional updates.
1014
1015
# Node Allocation (via Gossip)
1016
1017
New nodes must be allocated when a range is split. Instead of requiring
1018
every RoachNode to know about the status of all or even a large number
1019
of peer nodes --or-- alternatively requiring a specialized curator or
1020
master with sufficiently global knowledge, we use a gossip protocol to
1021
efficiently communicate only interesting information between all of the
1022
nodes in the cluster. What’s interesting information? One example would
1023
be whether a particular node has a lot of spare capacity. Each node,
1024
when gossiping, compares each topic of gossip to its own state. If its
1025
own state is somehow “more interesting” than the least interesting item
1026
in the topic it’s seen recently, it includes its own state as part of
1027
the next gossip session with a peer node. In this way, a node with
1028
capacity sufficiently in excess of the mean quickly becomes discovered
1029
by the entire cluster. To avoid piling onto outliers, nodes from the
1030
high capacity set are selected at random for allocation.
1031
1032
The gossip protocol itself contains two primary components:
1033
1034
- **Peer Selection**: each node maintains up to N peers with which it
1035
regularly communicates. It selects peers with an eye towards
1036
maximizing fanout. A peer node which itself communicates with an
1037
array of otherwise unknown nodes will be selected over one which
1038
communicates with a set containing significant overlap. Each time
1039
gossip is initiated, each nodes’ set of peers is exchanged. Each
1040
node is then free to incorporate the other’s peers as it sees fit.
1041
To avoid any node suffering from excess incoming requests, a node
1042
may refuse to answer a gossip exchange. Each node is biased
1043
towards answering requests from nodes without significant overlap
1044
and refusing requests otherwise.
1045
1046
Peers are efficiently selected using a heuristic as described in
1047
[Agarwal & Trachtenberg (2006)](https://drive.google.com/file/d/0B9GCVTp_FHJISmFRTThkOEZSM1U/edit?usp=sharing).
1048
1049
**TBD**: how to avoid partitions? Need to work out a simulation of
1050
the protocol to tune the behavior and see empirically how well it
1051
works.
1052
1053
- **Gossip Selection**: what to communicate. Gossip is divided into
1054
topics. Load characteristics (capacity per disk, cpu load, and
1055
state [e.g. draining, ok, failure]) are used to drive node
1056
allocation. Range statistics (range read/write load, missing
1057
replicas, unavailable ranges) and network topology (inter-rack
1058
bandwidth/latency, inter-datacenter bandwidth/latency, subnet
1059
outages) are used for determining when to split ranges, when to
1060
recover replicas vs. wait for network connectivity, and for
1061
debugging / sysops. In all cases, a set of minimums and a set of
1062
maximums is propagated; each node applies its own view of the
1063
world to augment the values. Each minimum and maximum value is
1064
tagged with the reporting node and other accompanying contextual
1065
information. Each topic of gossip has its own protobuf to hold the
1066
structured data. The number of items of gossip in each topic is
1067
limited by a configurable bound.
1068
1069
For efficiency, nodes assign each new item of gossip a sequence
1070
number and keep track of the highest sequence number each peer
1071
node has seen. Each round of gossip communicates only the delta
1072
containing new items.
1073
1074
# Node Accounting
1075
1076
The gossip protocol discussed in the previous section is useful to
1077
quickly communicate fragments of important information in a
1078
decentralized manner. However, complete accounting for each node is also
1079
stored to a central location, available to any dashboard process. This
1080
is done using the map itself. Each node periodically writes its state to
1081
the map with keys prefixed by `\0node`, similar to the first level of
1082
range metadata, but with an ‘`node`’ suffix. Each value is a protobuf
1083
containing the full complement of node statistics--everything
1084
communicated normally via the gossip protocol plus other useful, but
1085
non-critical data.
1086
1087
The range containing the first key in the node accounting table is
1088
responsible for gossiping the total count of nodes. This total count is
1089
used by the gossip network to most efficiently organize itself. In
1090
particular, the maximum number of hops for gossipped information to take
1091
before reaching a node is given by `ceil(log(node count) / log(max
1092
fanout)) + 1`.
1093
1094
# Key-prefix Accounting, Zones & Permissions
1095
1096
Arbitrarily fine-grained accounting and permissions are specified via
1097
key prefixes. Key prefixes can overlap, as is necessary for capturing
1098
hierarchical relationships. For illustrative purposes, let’s say keys
1099
specifying rows in a set of databases have the following format:
1100
1101
`<db>:<table>:<primary-key>[:<secondary-key>]`
1102
1103
In this case, we might collect accounting or specify permissions with
1104
key prefixes:
1105
1106
`db1`, `db1:user`, `db1:order`,
1107
1108
Accounting is kept for the entire map by default.
1109
1110
## Accounting
1111
to keep accounting for a range defined by a key prefix, an entry is created in
1112
the accounting system table. The format of accounting table keys is:
1113
1114
`\0acct<key-prefix>`
1115
1116
In practice, we assume each RoachNode capable of caching the
1117
entire accounting table as it is likely to be relatively small.
1118
1119
Accounting is kept for key prefix ranges with eventual consistency
1120
for efficiency. Updates to accounting values propagate through the
1121
system using the message queue facility if the accounting keys do
1122
not reside on the same range as ongoing activity (true for all but
1123
the smallest systems). There are two types of values which
1124
comprise accounting: counts and occurrences, for lack of better
1125
terms. Counts describe system state, such as the total number of
1126
bytes, rows, etc. Occurrences include transient performance and
1127
load metrics. Both types of accounting are captured as time series
1128
with minute granularity. The length of time accounting metrics are
1129
kept is configurable. Below are examples of each type of
1130
accounting value.
1131
1132
**System State Counters/Performance**
1133
1134
- Count of items (e.g. rows)
1135
- Total bytes
1136
- Total key bytes
1137
- Total value length
1138
- Queued message count
1139
- Queued message total bytes
1140
- Count of values \< 16B
1141
- Count of values \< 64B
1142
- Count of values \< 256B
1143
- Count of values \< 1K
1144
- Count of values \< 4K
1145
- Count of values \< 16K
1146
- Count of values \< 64K
1147
- Count of values \< 256K
1148
- Count of values \< 1M
1149
- Count of values \> 1M
1150
- Total bytes of accounting
1151
1152
1153
**Load Occurences**
1154
1155
Get op count
1156
Get total MB
1157
Put op count
1158
Put total MB
1159
Delete op count
1160
Delete total MB
1161
Delete range op count
1162
Delete range total MB
1163
Scan op count
1164
Scan op MB
1165
Split count
1166
Merge count
1167
1168
Because accounting information is kept as time series and over many
1169
possible metrics of interest, the data can become numerous. Accounting
1170
data are stored in the map near the key prefix described, in order to
1171
distribute load (for both aggregation and storage).
1172
1173
Accounting keys for system state have the form:
1174
`<key-prefix>|acctd<metric-name>*`. Notice the leading ‘pipe’
1175
character. It’s meant to sort the root level account AFTER any other
1176
system tables. They must increment the same underlying values as they
1177
are permanent counts, and not transient activity. Logic at the
1178
RoachNode takes care of snapshotting the value into an appropriately
1179
suffixed (e.g. with timestamp hour) multi-value time series entry.
1180
1181
Keys for perf/load metrics:
1182
`<key-prefix>acctd<metric-name><hourly-timestamp>`.
1183
1184
`<hourly-timestamp>`-suffixed accounting entries are multi-valued,
1185
containing a varint64 entry for each minute with activity during the
1186
specified hour.
1187
1188
To efficiently keep accounting over large key ranges, the task of
1189
aggregation must be distributed. If activity occurs within the same
1190
range as the key prefix for accounting, the updates are made as part
1191
of the consensus write. If the ranges differ, then a message is sent
1192
to the parent range to increment the accounting. If upon receiving the
1193
message, the parent range also does not include the key prefix, it in
1194
turn forwards it to its parent or left child in the balanced binary
1195
tree which is maintained to describe the range hierarchy. This limits
1196
the number of messages before an update is visible at the root to `2*log N`,
1197
where `N` is the number of ranges in the key prefix.
1198
1199
## Zones
1200
zones are stored in the map with keys prefixed by
1201
`\0zone` followed by the key prefix to which the zone
1202
configuration applies. Zone values specify a protobuf containing
1203
the datacenters from which replicas for ranges which fall under
1204
the zone must be chosen.
1205
1206
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`.
1207
1208
If zones are modified in situ, each RoachNode verifies the
1209
existing zones for its ranges against the zone configuration. If
1210
it discovers differences, it reconfigures ranges in the same way
1211
that it rebalances away from busy nodes, via special-case 1:1
1212
split to a duplicate range comprising the new configuration.
1213
1214
### Permissions
1215
permissions are stored in the map with keys prefixed by *\0perm* followed by
1216
the key prefix and user to which the specified permissions apply. The format of
1217
permissions keys is:
1218
1219
`\0perm<key-prefix><user>`
1220
1221
Permission values are a protobuf containing the permission details;
1222
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`.
1223
1224
A default system root permission is assumed for the entire map
1225
with an empty key prefix and read/write as true.
1226
1227
When determining whether or not to allow a read or a write a key
1228
value (e.g. `db1:user:1` for user `spencer`), a RoachNode would
1229
read the following permissions values:
1230
1231
```
1232
\0perm<db1:user:1>spencer
1233
\0perm<db1:user>spencer
1234
\0perm<db1>spencer
1235
\0perm<>spencer
1236
```
1237
1238
If any prefix in the hierarchy provides the required permission,
1239
the request is satisfied; otherwise, the request returns an
1240
error.
1241
1242
The priority for a user permission is used to order requests at
1243
Raft consensus ranges and for choosing an initial priority for
1244
distributed transactions. When scheduling operations at the Raft
1245
consensus range, all outstanding requests are ordered by key
1246
prefix and each assigned priorities according to key, user and
1247
arrival time. The next request is chosen probabilistically using
1248
priorities to weight the choice. Each key can have multiple
1249
priorities as they’re hierarchical (e.g. for /user/key, one
1250
priority for root ‘/’, and one for ‘/user/key’). The most general
1251
priority is used first. If two keys share the most general, then
1252
they’re compared with the next most general if applicable, and so on.
1253
1254
# Key-Value API
1255
1256
see the protobufs in [proto/](https://github.com/cockroachdb/cockroach/blob/master/proto),
1257
in particular [proto/api.proto](https://github.com/cockroachdb/cockroach/blob/master/proto/api.proto) and the comments within.
1258
1259
# Structured Data API
1260
1261
A preliminary design can be found in the [Go source documentation](http://godoc.org/github.com/cockroachdb/cockroach/structured).