Line data Source code
1 : //! See `pageserver_api::shard` for description on sharding.
2 :
3 : use std::{ops::RangeInclusive, str::FromStr};
4 :
5 : use hex::FromHex;
6 : use serde::{Deserialize, Serialize};
7 :
8 : use crate::id::TenantId;
9 :
10 0 : #[derive(Ord, PartialOrd, Eq, PartialEq, Clone, Copy, Serialize, Deserialize, Debug, Hash)]
11 : pub struct ShardNumber(pub u8);
12 :
13 0 : #[derive(Ord, PartialOrd, Eq, PartialEq, Clone, Copy, Serialize, Deserialize, Debug, Hash)]
14 : pub struct ShardCount(pub u8);
15 :
16 : /// Combination of ShardNumber and ShardCount.
17 : ///
18 : /// For use within the context of a particular tenant, when we need to know which shard we're
19 : /// dealing with, but do not need to know the full ShardIdentity (because we won't be doing
20 : /// any page->shard mapping), and do not need to know the fully qualified TenantShardId.
21 : #[derive(Eq, PartialEq, PartialOrd, Ord, Clone, Copy, Hash)]
22 : pub struct ShardIndex {
23 : pub shard_number: ShardNumber,
24 : pub shard_count: ShardCount,
25 : }
26 :
27 : /// Formatting helper, for generating the `shard_id` label in traces.
28 : pub struct ShardSlug<'a>(&'a TenantShardId);
29 :
30 : /// TenantShardId globally identifies a particular shard in a particular tenant.
31 : ///
32 : /// These are written as `<TenantId>-<ShardSlug>`, for example:
33 : /// # The second shard in a two-shard tenant
34 : /// 072f1291a5310026820b2fe4b2968934-0102
35 : ///
36 : /// If the `ShardCount` is _unsharded_, the `TenantShardId` is written without
37 : /// a shard suffix and is equivalent to the encoding of a `TenantId`: this enables
38 : /// an unsharded [`TenantShardId`] to be used interchangably with a [`TenantId`].
39 : ///
40 : /// The human-readable encoding of an unsharded TenantShardId, such as used in API URLs,
41 : /// is both forward and backward compatible with TenantId: a legacy TenantId can be
42 : /// decoded as a TenantShardId, and when re-encoded it will be parseable
43 : /// as a TenantId.
44 : #[derive(Eq, PartialEq, PartialOrd, Ord, Clone, Copy, Hash)]
45 : pub struct TenantShardId {
46 : pub tenant_id: TenantId,
47 : pub shard_number: ShardNumber,
48 : pub shard_count: ShardCount,
49 : }
50 :
51 : impl ShardCount {
52 : pub const MAX: Self = Self(u8::MAX);
53 : pub const MIN: Self = Self(0);
54 :
55 : /// The internal value of a ShardCount may be zero, which means "1 shard, but use
56 : /// legacy format for TenantShardId that excludes the shard suffix", also known
57 : /// as [`TenantShardId::unsharded`].
58 : ///
59 : /// This method returns the actual number of shards, i.e. if our internal value is
60 : /// zero, we return 1 (unsharded tenants have 1 shard).
61 4809234 : pub fn count(&self) -> u8 {
62 4809234 : if self.0 > 0 {
63 5018 : self.0
64 : } else {
65 4804216 : 1
66 : }
67 4809234 : }
68 :
69 : /// The literal internal value: this is **not** the number of shards in the
70 : /// tenant, as we have a special zero value for legacy unsharded tenants. Use
71 : /// [`Self::count`] if you want to know the cardinality of shards.
72 2 : pub fn literal(&self) -> u8 {
73 2 : self.0
74 2 : }
75 :
76 : /// Whether the `ShardCount` is for an unsharded tenant, so uses one shard but
77 : /// uses the legacy format for `TenantShardId`. See also the documentation for
78 : /// [`Self::count`].
79 0 : pub fn is_unsharded(&self) -> bool {
80 0 : self.0 == 0
81 0 : }
82 :
83 : /// `v` may be zero, or the number of shards in the tenant. `v` is what
84 : /// [`Self::literal`] would return.
85 9027 : pub const fn new(val: u8) -> Self {
86 9027 : Self(val)
87 9027 : }
88 : }
89 :
90 : impl ShardNumber {
91 : pub const MAX: Self = Self(u8::MAX);
92 : }
93 :
94 : impl TenantShardId {
95 39 : pub fn unsharded(tenant_id: TenantId) -> Self {
96 39 : Self {
97 39 : tenant_id,
98 39 : shard_number: ShardNumber(0),
99 39 : shard_count: ShardCount(0),
100 39 : }
101 39 : }
102 :
103 : /// The range of all TenantShardId that belong to a particular TenantId. This is useful when
104 : /// you have a BTreeMap of TenantShardId, and are querying by TenantId.
105 0 : pub fn tenant_range(tenant_id: TenantId) -> RangeInclusive<Self> {
106 0 : RangeInclusive::new(
107 0 : Self {
108 0 : tenant_id,
109 0 : shard_number: ShardNumber(0),
110 0 : shard_count: ShardCount(0),
111 0 : },
112 0 : Self {
113 0 : tenant_id,
114 0 : shard_number: ShardNumber::MAX,
115 0 : shard_count: ShardCount::MAX,
116 0 : },
117 0 : )
118 0 : }
119 :
120 24521 : pub fn shard_slug(&self) -> impl std::fmt::Display + '_ {
121 24521 : ShardSlug(self)
122 24521 : }
123 :
124 : /// Convenience for code that has special behavior on the 0th shard.
125 516 : pub fn is_shard_zero(&self) -> bool {
126 516 : self.shard_number == ShardNumber(0)
127 516 : }
128 :
129 : /// The "unsharded" value is distinct from simply having a single shard: it represents
130 : /// a tenant which is not shard-aware at all, and whose storage paths will not include
131 : /// a shard suffix.
132 0 : pub fn is_unsharded(&self) -> bool {
133 0 : self.shard_number == ShardNumber(0) && self.shard_count.is_unsharded()
134 0 : }
135 :
136 : /// Convenience for dropping the tenant_id and just getting the ShardIndex: this
137 : /// is useful when logging from code that is already in a span that includes tenant ID, to
138 : /// keep messages reasonably terse.
139 0 : pub fn to_index(&self) -> ShardIndex {
140 0 : ShardIndex {
141 0 : shard_number: self.shard_number,
142 0 : shard_count: self.shard_count,
143 0 : }
144 0 : }
145 :
146 : /// Calculate the children of this TenantShardId when splitting the overall tenant into
147 : /// the given number of shards.
148 6 : pub fn split(&self, new_shard_count: ShardCount) -> Vec<TenantShardId> {
149 6 : let effective_old_shard_count = std::cmp::max(self.shard_count.0, 1);
150 6 : let mut child_shards = Vec::new();
151 32 : for shard_number in 0..ShardNumber(new_shard_count.0).0 {
152 : // Key mapping is based on a round robin mapping of key hash modulo shard count,
153 : // so our child shards are the ones which the same keys would map to.
154 32 : if shard_number % effective_old_shard_count == self.shard_number.0 {
155 28 : child_shards.push(TenantShardId {
156 28 : tenant_id: self.tenant_id,
157 28 : shard_number: ShardNumber(shard_number),
158 28 : shard_count: new_shard_count,
159 28 : })
160 4 : }
161 : }
162 :
163 6 : child_shards
164 6 : }
165 : }
166 :
167 : impl std::fmt::Display for ShardNumber {
168 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
169 0 : self.0.fmt(f)
170 0 : }
171 : }
172 :
173 : impl std::fmt::Display for ShardSlug<'_> {
174 11956 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
175 11956 : write!(
176 11956 : f,
177 11956 : "{:02x}{:02x}",
178 11956 : self.0.shard_number.0, self.0.shard_count.0
179 11956 : )
180 11956 : }
181 : }
182 :
183 : impl std::fmt::Display for TenantShardId {
184 13430 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
185 13430 : if self.shard_count != ShardCount(0) {
186 153 : write!(f, "{}-{}", self.tenant_id, self.shard_slug())
187 : } else {
188 : // Legacy case (shard_count == 0) -- format as just the tenant id. Note that this
189 : // is distinct from the normal single shard case (shard count == 1).
190 13277 : self.tenant_id.fmt(f)
191 : }
192 13430 : }
193 : }
194 :
195 : impl std::fmt::Debug for TenantShardId {
196 3132 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
197 3132 : // Debug is the same as Display: the compact hex representation
198 3132 : write!(f, "{}", self)
199 3132 : }
200 : }
201 :
202 : impl std::str::FromStr for TenantShardId {
203 : type Err = hex::FromHexError;
204 :
205 4537 : fn from_str(s: &str) -> Result<Self, Self::Err> {
206 4537 : // Expect format: 16 byte TenantId, '-', 1 byte shard number, 1 byte shard count
207 4537 : if s.len() == 32 {
208 : // Legacy case: no shard specified
209 : Ok(Self {
210 4501 : tenant_id: TenantId::from_str(s)?,
211 4501 : shard_number: ShardNumber(0),
212 4501 : shard_count: ShardCount(0),
213 : })
214 36 : } else if s.len() == 37 {
215 36 : let bytes = s.as_bytes();
216 36 : let tenant_id = TenantId::from_hex(&bytes[0..32])?;
217 36 : let mut shard_parts: [u8; 2] = [0u8; 2];
218 36 : hex::decode_to_slice(&bytes[33..37], &mut shard_parts)?;
219 36 : Ok(Self {
220 36 : tenant_id,
221 36 : shard_number: ShardNumber(shard_parts[0]),
222 36 : shard_count: ShardCount(shard_parts[1]),
223 36 : })
224 : } else {
225 0 : Err(hex::FromHexError::InvalidStringLength)
226 : }
227 4537 : }
228 : }
229 :
230 : impl From<[u8; 18]> for TenantShardId {
231 2 : fn from(b: [u8; 18]) -> Self {
232 2 : let tenant_id_bytes: [u8; 16] = b[0..16].try_into().unwrap();
233 2 :
234 2 : Self {
235 2 : tenant_id: TenantId::from(tenant_id_bytes),
236 2 : shard_number: ShardNumber(b[16]),
237 2 : shard_count: ShardCount(b[17]),
238 2 : }
239 2 : }
240 : }
241 :
242 : impl ShardIndex {
243 0 : pub fn new(number: ShardNumber, count: ShardCount) -> Self {
244 0 : Self {
245 0 : shard_number: number,
246 0 : shard_count: count,
247 0 : }
248 0 : }
249 110 : pub fn unsharded() -> Self {
250 110 : Self {
251 110 : shard_number: ShardNumber(0),
252 110 : shard_count: ShardCount(0),
253 110 : }
254 110 : }
255 :
256 : /// The "unsharded" value is distinct from simply having a single shard: it represents
257 : /// a tenant which is not shard-aware at all, and whose storage paths will not include
258 : /// a shard suffix.
259 72927 : pub fn is_unsharded(&self) -> bool {
260 72927 : self.shard_number == ShardNumber(0) && self.shard_count == ShardCount(0)
261 72927 : }
262 :
263 : /// For use in constructing remote storage paths: concatenate this with a TenantId
264 : /// to get a fully qualified TenantShardId.
265 : ///
266 : /// Backward compat: this function returns an empty string if Self::is_unsharded, such
267 : /// that the legacy pre-sharding remote key format is preserved.
268 1723 : pub fn get_suffix(&self) -> String {
269 1723 : if self.is_unsharded() {
270 1715 : "".to_string()
271 : } else {
272 8 : format!("-{:02x}{:02x}", self.shard_number.0, self.shard_count.0)
273 : }
274 1723 : }
275 : }
276 :
277 : impl std::fmt::Display for ShardIndex {
278 2084 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
279 2084 : write!(f, "{:02x}{:02x}", self.shard_number.0, self.shard_count.0)
280 2084 : }
281 : }
282 :
283 : impl std::fmt::Debug for ShardIndex {
284 1556 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
285 1556 : // Debug is the same as Display: the compact hex representation
286 1556 : write!(f, "{}", self)
287 1556 : }
288 : }
289 :
290 : impl std::str::FromStr for ShardIndex {
291 : type Err = hex::FromHexError;
292 :
293 3129 : fn from_str(s: &str) -> Result<Self, Self::Err> {
294 3129 : // Expect format: 1 byte shard number, 1 byte shard count
295 3129 : if s.len() == 4 {
296 3129 : let bytes = s.as_bytes();
297 3129 : let mut shard_parts: [u8; 2] = [0u8; 2];
298 3129 : hex::decode_to_slice(bytes, &mut shard_parts)?;
299 3129 : Ok(Self {
300 3129 : shard_number: ShardNumber(shard_parts[0]),
301 3129 : shard_count: ShardCount(shard_parts[1]),
302 3129 : })
303 : } else {
304 0 : Err(hex::FromHexError::InvalidStringLength)
305 : }
306 3129 : }
307 : }
308 :
309 : impl From<[u8; 2]> for ShardIndex {
310 1 : fn from(b: [u8; 2]) -> Self {
311 1 : Self {
312 1 : shard_number: ShardNumber(b[0]),
313 1 : shard_count: ShardCount(b[1]),
314 1 : }
315 1 : }
316 : }
317 :
318 : impl Serialize for TenantShardId {
319 46 : fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
320 46 : where
321 46 : S: serde::Serializer,
322 46 : {
323 46 : if serializer.is_human_readable() {
324 42 : serializer.collect_str(self)
325 : } else {
326 : // Note: while human encoding of [`TenantShardId`] is backward and forward
327 : // compatible, this binary encoding is not.
328 4 : let mut packed: [u8; 18] = [0; 18];
329 4 : packed[0..16].clone_from_slice(&self.tenant_id.as_arr());
330 4 : packed[16] = self.shard_number.0;
331 4 : packed[17] = self.shard_count.0;
332 4 :
333 4 : packed.serialize(serializer)
334 : }
335 46 : }
336 : }
337 :
338 : impl<'de> Deserialize<'de> for TenantShardId {
339 9 : fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
340 9 : where
341 9 : D: serde::Deserializer<'de>,
342 9 : {
343 : struct IdVisitor {
344 : is_human_readable_deserializer: bool,
345 : }
346 :
347 : impl<'de> serde::de::Visitor<'de> for IdVisitor {
348 : type Value = TenantShardId;
349 :
350 0 : fn expecting(&self, formatter: &mut std::fmt::Formatter) -> std::fmt::Result {
351 0 : if self.is_human_readable_deserializer {
352 0 : formatter.write_str("value in form of hex string")
353 : } else {
354 0 : formatter.write_str("value in form of integer array([u8; 18])")
355 : }
356 0 : }
357 :
358 2 : fn visit_seq<A>(self, seq: A) -> Result<Self::Value, A::Error>
359 2 : where
360 2 : A: serde::de::SeqAccess<'de>,
361 2 : {
362 2 : let s = serde::de::value::SeqAccessDeserializer::new(seq);
363 2 : let id: [u8; 18] = Deserialize::deserialize(s)?;
364 2 : Ok(TenantShardId::from(id))
365 2 : }
366 :
367 7 : fn visit_str<E>(self, v: &str) -> Result<Self::Value, E>
368 7 : where
369 7 : E: serde::de::Error,
370 7 : {
371 7 : TenantShardId::from_str(v).map_err(E::custom)
372 7 : }
373 : }
374 :
375 9 : if deserializer.is_human_readable() {
376 7 : deserializer.deserialize_str(IdVisitor {
377 7 : is_human_readable_deserializer: true,
378 7 : })
379 : } else {
380 2 : deserializer.deserialize_tuple(
381 2 : 18,
382 2 : IdVisitor {
383 2 : is_human_readable_deserializer: false,
384 2 : },
385 2 : )
386 : }
387 9 : }
388 : }
389 :
390 : impl Serialize for ShardIndex {
391 18 : fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
392 18 : where
393 18 : S: serde::Serializer,
394 18 : {
395 18 : if serializer.is_human_readable() {
396 16 : serializer.collect_str(self)
397 : } else {
398 : // Binary encoding is not used in index_part.json, but is included in anticipation of
399 : // switching various structures (e.g. inter-process communication, remote metadata) to more
400 : // compact binary encodings in future.
401 2 : let mut packed: [u8; 2] = [0; 2];
402 2 : packed[0] = self.shard_number.0;
403 2 : packed[1] = self.shard_count.0;
404 2 : packed.serialize(serializer)
405 : }
406 18 : }
407 : }
408 :
409 : impl<'de> Deserialize<'de> for ShardIndex {
410 3129 : fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
411 3129 : where
412 3129 : D: serde::Deserializer<'de>,
413 3129 : {
414 : struct IdVisitor {
415 : is_human_readable_deserializer: bool,
416 : }
417 :
418 : impl<'de> serde::de::Visitor<'de> for IdVisitor {
419 : type Value = ShardIndex;
420 :
421 0 : fn expecting(&self, formatter: &mut std::fmt::Formatter) -> std::fmt::Result {
422 0 : if self.is_human_readable_deserializer {
423 0 : formatter.write_str("value in form of hex string")
424 : } else {
425 0 : formatter.write_str("value in form of integer array([u8; 2])")
426 : }
427 0 : }
428 :
429 1 : fn visit_seq<A>(self, seq: A) -> Result<Self::Value, A::Error>
430 1 : where
431 1 : A: serde::de::SeqAccess<'de>,
432 1 : {
433 1 : let s = serde::de::value::SeqAccessDeserializer::new(seq);
434 1 : let id: [u8; 2] = Deserialize::deserialize(s)?;
435 1 : Ok(ShardIndex::from(id))
436 1 : }
437 :
438 3128 : fn visit_str<E>(self, v: &str) -> Result<Self::Value, E>
439 3128 : where
440 3128 : E: serde::de::Error,
441 3128 : {
442 3128 : ShardIndex::from_str(v).map_err(E::custom)
443 3128 : }
444 : }
445 :
446 3129 : if deserializer.is_human_readable() {
447 3128 : deserializer.deserialize_str(IdVisitor {
448 3128 : is_human_readable_deserializer: true,
449 3128 : })
450 : } else {
451 1 : deserializer.deserialize_tuple(
452 1 : 2,
453 1 : IdVisitor {
454 1 : is_human_readable_deserializer: false,
455 1 : },
456 1 : )
457 : }
458 3129 : }
459 : }
|
