mz_persist_client/
cfg.rs

1// Copyright Materialize, Inc. and contributors. All rights reserved.
2//
3// Use of this software is governed by the Business Source License
4// included in the LICENSE file.
5//
6// As of the Change Date specified in that file, in accordance with
7// the Business Source License, use of this software will be governed
8// by the Apache License, Version 2.0.
9
10#![allow(missing_docs)]
11
12//! The tunable knobs for persist.
13//!
14//! Persist configs are `ParameterScope::Environment` as a class, including the
15//! ones only ever read on `clusterd`. The same client code runs in
16//! `environmentd`, and every copy of it acts on shared durable state, so a
17//! replica-scoped persist config could not reach the `environmentd` client and
18//! a rollout targeting replicas would leave a shard's other writer on the old
19//! value indefinitely. Declare new persist configs `Environment` too.
20
21use std::sync::Arc;
22use std::sync::atomic::AtomicBool;
23use std::time::{Duration, SystemTime, UNIX_EPOCH};
24
25use mz_build_info::BuildInfo;
26use mz_dyncfg::{Config, ConfigDefault, ConfigSet, ConfigUpdates, ParameterScope};
27use mz_ore::instrument;
28use mz_ore::now::NowFn;
29use mz_persist::cfg::BlobKnobs;
30use mz_persist::retry::Retry;
31use mz_persist::s3::{BLOB_HTTP_POOL_MAX_IN_FLIGHT_PER_CLIENT, BlobHttpPool};
32use mz_postgres_client::PostgresClientKnobs;
33use proptest_derive::Arbitrary;
34use semver::Version;
35use serde::{Deserialize, Serialize};
36use tokio::sync::watch;
37
38use crate::async_runtime;
39use crate::internal::machine::{
40    NEXT_LISTEN_BATCH_RETRYER_CLAMP, NEXT_LISTEN_BATCH_RETRYER_INITIAL_BACKOFF,
41    NEXT_LISTEN_BATCH_RETRYER_MULTIPLIER,
42};
43use crate::internal::state::ROLLUP_THRESHOLD;
44use crate::operators::STORAGE_SOURCE_DECODE_FUEL;
45use crate::read::READER_LEASE_DURATION;
46
47// Ignores the patch version
48const SELF_MANAGED_VERSIONS: &[Version; 2] = &[
49    // 25.1
50    Version::new(0, 130, 0),
51    // 25.2
52    Version::new(0, 147, 0),
53];
54
55/// The tunable knobs for persist.
56///
57/// Tuning inputs:
58/// - A larger blob_target_size (capped at KEY_VAL_DATA_MAX_LEN) results in
59///   fewer entries in consensus state. Before we have compaction and/or
60///   incremental state, it is already growing without bound, so this is a
61///   concern. OTOH, for any "reasonable" size (> 100MiB?) of blob_target_size,
62///   it seems we'd end up with a pretty tremendous amount of data in the shard
63///   before this became a real issue.
64/// - A larger blob_target_size will results in fewer s3 operations, which are
65///   charged per operation. (Hmm, maybe not if we're charged per call in a
66///   multipart op. The S3Blob impl already chunks things at 8MiB.)
67/// - A smaller blob_target_size will result in more even memory usage in
68///   readers.
69/// - A larger batch_builder_max_outstanding_parts increases throughput (to a
70///   point).
71/// - A smaller batch_builder_max_outstanding_parts provides a bound on the
72///   amount of memory used by a writer.
73/// - A larger compaction_heuristic_min_inputs means state size is larger.
74/// - A smaller compaction_heuristic_min_inputs means more compactions happen
75///   (higher write amp).
76/// - A larger compaction_heuristic_min_updates means more consolidations are
77///   discovered while reading a snapshot (higher read amp and higher space
78///   amp).
79/// - A smaller compaction_heuristic_min_updates means more compactions happen
80///   (higher write amp).
81///
82/// Tuning logic:
83/// - blob_target_size was initially selected to be an exact multiple of 8MiB
84///   (the s3 multipart size) that was in the same neighborhood as our initial
85///   max throughput (~250MiB).
86/// - batch_builder_max_outstanding_parts was initially selected to be as small
87///   as possible without harming pipelining. 0 means no pipelining, 1 is full
88///   pipelining as long as generating data takes less time than writing to s3
89///   (hopefully a fair assumption), 2 is a little extra slop on top of 1.
90/// - compaction_heuristic_min_inputs was set by running the open-loop benchmark
91///   with batches of size 10,240 bytes (selected to be small but such that the
92///   overhead of our columnar encoding format was less than 10%) and manually
93///   increased until the write amp stopped going down. This becomes much less
94///   important once we have incremental state. The initial value is a
95///   placeholder and should be revisited at some point.
96/// - compaction_heuristic_min_updates was set via a thought experiment. This is
97///   an `O(n*log(n))` upper bound on the number of unconsolidated updates that
98///   would be consolidated if we compacted as the in-mem Spine does. The
99///   initial value is a placeholder and should be revisited at some point.
100///
101/// TODO: Move these tuning notes into SessionVar descriptions once we have
102/// SystemVars for most of these.
103//
104// TODO: The configs left here don't react dynamically to changes. Move as many
105// of them to DynamicConfig as possible.
106#[derive(Debug, Clone)]
107pub struct PersistConfig {
108    /// Info about which version of the code is running.
109    pub build_version: Version,
110    /// An opaque string describing the host of this persist client.
111    /// Stored in state and used for debugging.
112    pub hostname: String,
113    /// Whether this persist instance is running in a "cc" sized cluster.
114    pub is_cc_active: bool,
115    /// Memory limit of the process, if known.
116    pub announce_memory_limit: Option<usize>,
117    /// A clock to use for all leasing and other non-debugging use.
118    pub now: NowFn,
119    /// Persist [Config]s that can change value dynamically within the lifetime
120    /// of a process.
121    ///
122    /// TODO(cfg): Entirely replace dynamic with this.
123    pub configs: Arc<ConfigSet>,
124    /// Indicates whether `configs` has been synced at least once with an
125    /// upstream source.
126    configs_synced_once: Arc<watch::Sender<bool>>,
127    /// Whether to physically and logically compact batches in blob storage.
128    pub compaction_enabled: bool,
129    /// Whether the `Compactor` will process compaction requests, or drop them on the floor.
130    pub compaction_process_requests: Arc<AtomicBool>,
131    /// In Compactor::compact_and_apply_background, the maximum number of concurrent
132    /// compaction requests that can execute for a given shard.
133    pub compaction_concurrency_limit: usize,
134    /// In Compactor::compact_and_apply_background, the maximum number of pending
135    /// compaction requests to queue.
136    pub compaction_queue_size: usize,
137    /// In Compactor::compact_and_apply_background, how many updates to encode or
138    /// decode before voluntarily yielding the task.
139    pub compaction_yield_after_n_updates: usize,
140    /// Length of time after a writer's last operation after which the writer
141    /// may be expired.
142    pub writer_lease_duration: Duration,
143    /// Length of time between critical handles' calls to downgrade since
144    pub critical_downgrade_interval: Duration,
145    /// Number of worker threads to create for the [`crate::IsolatedRuntime`], defaults to the
146    /// number of threads.
147    pub isolated_runtime_worker_threads: usize,
148}
149
150// Impl Deref to ConfigSet for convenience of accessing the dynamic configs.
151impl std::ops::Deref for PersistConfig {
152    type Target = ConfigSet;
153    fn deref(&self) -> &Self::Target {
154        &self.configs
155    }
156}
157
158impl PersistConfig {
159    /// Returns a new instance of [PersistConfig] with default tuning and
160    /// default ConfigSet.
161    pub fn new_default_configs(build_info: &BuildInfo, now: NowFn) -> Self {
162        Self::new(build_info, now, all_dyncfgs(ConfigSet::default()))
163    }
164
165    /// Returns a new instance of [PersistConfig] with default tuning and the
166    /// specified ConfigSet.
167    pub fn new(build_info: &BuildInfo, now: NowFn, configs: ConfigSet) -> Self {
168        // Escape hatch in case we need to disable compaction.
169        let compaction_disabled = mz_ore::env::is_var_truthy("MZ_PERSIST_COMPACTION_DISABLED");
170
171        // We create receivers on demand, so we drop the initial receiver.
172        let (configs_synced_once, _) = watch::channel(false);
173
174        Self {
175            build_version: build_info.semver_version(),
176            is_cc_active: false,
177            announce_memory_limit: None,
178            now,
179            configs: Arc::new(configs),
180            configs_synced_once: Arc::new(configs_synced_once),
181            compaction_enabled: !compaction_disabled,
182            compaction_process_requests: Arc::new(AtomicBool::new(true)),
183            compaction_concurrency_limit: 5,
184            compaction_queue_size: 20,
185            compaction_yield_after_n_updates: 100_000,
186            writer_lease_duration: 60 * Duration::from_secs(60),
187            critical_downgrade_interval: Duration::from_secs(30),
188            isolated_runtime_worker_threads: num_cpus::get(),
189            // TODO: This doesn't work with the process orchestrator. Instead,
190            // separate --log-prefix into --service-name and --enable-log-prefix
191            // options, where the first is always provided and the second is
192            // conditionally enabled by the process orchestrator.
193            hostname: {
194                use std::fmt::Write;
195                let mut name = std::env::var("HOSTNAME").unwrap_or_else(|_| "unknown".to_owned());
196                write!(&mut name, " {}", build_info.version)
197                    .expect("writing to string should not fail");
198                name
199            },
200        }
201    }
202
203    pub(crate) fn set_config<T: ConfigDefault>(&self, cfg: &Config<T>, val: T) {
204        let mut updates = ConfigUpdates::default();
205        updates.add(cfg, val);
206        updates.apply(self)
207    }
208
209    /// Applies the provided updates to this configuration.
210    ///
211    /// You should prefer calling this method over mutating `self.configs`
212    /// directly, so that [`Self::configs_synced_once`] can be properly
213    /// maintained.
214    pub fn apply_from(&self, updates: &ConfigUpdates) {
215        updates.apply(&self.configs);
216        self.configs_synced_once.send_replace(true);
217    }
218
219    /// Resolves when `configs` has been synced at least once with an upstream
220    /// source, i.e., via [`Self::apply_from`].
221    ///
222    /// If `configs` has already been synced once at the time the method is
223    /// called, resolves immediately.
224    ///
225    /// Useful in conjunction with configuration parameters that cannot be
226    /// dynamically updated once set (e.g., PubSub).
227    #[instrument(level = "info")]
228    pub async fn configs_synced_once(&self) {
229        self.configs_synced_once
230            .subscribe()
231            .wait_for(|synced| *synced)
232            .await
233            .expect("we have a borrow on sender so it cannot drop");
234    }
235
236    /// The maximum amount of work to do in the persist_source mfp_and_decode
237    /// operator before yielding.
238    pub fn storage_source_decode_fuel(&self) -> usize {
239        STORAGE_SOURCE_DECODE_FUEL.get(self)
240    }
241
242    /// Overrides the value for "persist_reader_lease_duration".
243    pub fn set_reader_lease_duration(&self, val: Duration) {
244        self.set_config(&READER_LEASE_DURATION, val);
245    }
246
247    /// Overrides the value for "persist_rollup_threshold".
248    pub fn set_rollup_threshold(&self, val: usize) {
249        self.set_config(&ROLLUP_THRESHOLD, val);
250    }
251
252    /// Overrides the value for the "persist_next_listen_batch_retryer_*"
253    /// configs.
254    pub fn set_next_listen_batch_retryer(&self, val: RetryParameters) {
255        self.set_config(
256            &NEXT_LISTEN_BATCH_RETRYER_INITIAL_BACKOFF,
257            val.initial_backoff,
258        );
259        self.set_config(&NEXT_LISTEN_BATCH_RETRYER_MULTIPLIER, val.multiplier);
260        self.set_config(&NEXT_LISTEN_BATCH_RETRYER_CLAMP, val.clamp);
261    }
262
263    pub fn disable_compaction(&self) {
264        tracing::info!("Disabling Persist Compaction");
265        self.compaction_process_requests
266            .store(false, std::sync::atomic::Ordering::Relaxed);
267    }
268
269    pub fn enable_compaction(&self) {
270        tracing::info!("Enabling Persist Compaction");
271        self.compaction_process_requests
272            .store(true, std::sync::atomic::Ordering::Relaxed);
273    }
274
275    /// Returns a new instance of [PersistConfig] for tests.
276    pub fn new_for_tests() -> Self {
277        use mz_build_info::DUMMY_BUILD_INFO;
278        use mz_ore::now::SYSTEM_TIME;
279
280        let mut cfg = Self::new_default_configs(&DUMMY_BUILD_INFO, SYSTEM_TIME.clone());
281        cfg.hostname = "tests".into();
282        cfg.isolated_runtime_worker_threads = async_runtime::TEST_THREADS;
283        cfg
284    }
285}
286
287#[allow(non_upper_case_globals)]
288pub(crate) const MiB: usize = 1024 * 1024;
289
290/// Adds the full set of all persist [Config]s.
291///
292/// Persist configs are [`ParameterScope::Environment`] by default, because a
293/// persist client runs in every `clusterd` process and the per-replica dyncfg
294/// push reaches its `ConfigSet` (the compute worker applies the pushed updates
295/// to `persist_clients.cfg()`). `environmentd`'s own persist clients read the
296/// same configs and see the environment-wide value, which is what "replica-
297/// local" means for a config that both processes run. The exceptions are the
298/// configs read only from `environmentd`'s catalog and expression cache, which
299/// are [`ParameterScope::Environment`].
300///
301/// TODO(cfg): Consider replacing this with a static global registry powered by
302/// something like the `ctor` or `inventory` crate. This would involve managing
303/// the footgun of a Config being linked into one binary but not the other.
304pub fn all_dyncfgs(configs: ConfigSet) -> ConfigSet {
305    mz_persist::cfg::all_dyn_configs(configs)
306        .add(&crate::batch::BATCH_DELETE_ENABLED)
307        .add(&crate::batch::BLOB_TARGET_SIZE)
308        .add(&crate::batch::INLINE_WRITES_TOTAL_MAX_BYTES)
309        .add(&crate::batch::INLINE_WRITES_SINGLE_MAX_BYTES)
310        .add(&crate::batch::ENCODING_ENABLE_DICTIONARY)
311        .add(&crate::batch::ENCODING_COMPRESSION_FORMAT)
312        .add(&crate::batch::STRUCTURED_KEY_LOWER_LEN)
313        .add(&crate::batch::MAX_RUN_LEN)
314        .add(&crate::batch::MAX_RUNS)
315        .add(&BLOB_OPERATION_TIMEOUT)
316        .add(&BLOB_OPERATION_ATTEMPT_TIMEOUT)
317        .add(&BLOB_CONNECT_TIMEOUT)
318        .add(&BLOB_READ_TIMEOUT)
319        .add(&SHARD_METRICS)
320        .add(&PER_SHARD_METRICS_ENABLE_REGEX)
321        .add(&SHARD_INFO_METRIC)
322        .add(&crate::cfg::CONSENSUS_CONNECTION_POOL_MAX_SIZE)
323        .add(&crate::cfg::CONSENSUS_CONNECTION_POOL_MAX_WAIT)
324        .add(&crate::cfg::CONSENSUS_CONNECTION_POOL_TTL_STAGGER)
325        .add(&crate::cfg::CONSENSUS_CONNECTION_POOL_TTL)
326        .add(&crate::cfg::CRDB_CONNECT_TIMEOUT)
327        .add(&crate::cfg::CRDB_TCP_USER_TIMEOUT)
328        .add(&crate::cfg::CRDB_KEEPALIVES_IDLE)
329        .add(&crate::cfg::CRDB_KEEPALIVES_INTERVAL)
330        .add(&crate::cfg::CRDB_KEEPALIVES_RETRIES)
331        .add(&crate::cfg::USE_CRITICAL_SINCE_TXN)
332        .add(&crate::cfg::USE_CRITICAL_SINCE_CATALOG)
333        .add(&crate::cfg::USE_CRITICAL_SINCE_SOURCE)
334        .add(&crate::cfg::USE_CRITICAL_SINCE_SNAPSHOT)
335        .add(&crate::cfg::SOURCE_HYDRATION_FRONTIER_COALESCE_BYTES)
336        .add(&crate::cfg::SOURCE_FETCH_CONCURRENCY)
337        .add(&BATCH_BUILDER_MAX_OUTSTANDING_PARTS)
338        .add(&COMPACTION_HEURISTIC_MIN_INPUTS)
339        .add(&COMPACTION_HEURISTIC_MIN_PARTS)
340        .add(&COMPACTION_HEURISTIC_MIN_UPDATES)
341        .add(&COMPACTION_MEMORY_BOUND_BYTES)
342        .add(&GC_BLOB_DELETE_CONCURRENCY_LIMIT)
343        .add(&STATE_VERSIONS_RECENT_LIVE_DIFFS_LIMIT)
344        .add(&USAGE_STATE_FETCH_CONCURRENCY_LIMIT)
345        .add(&crate::cache::STATE_UPDATE_LEASE_TIMEOUT)
346        .add(&crate::cli::admin::CATALOG_FORCE_COMPACTION_FUEL)
347        .add(&crate::cli::admin::CATALOG_FORCE_COMPACTION_WAIT)
348        .add(&crate::cli::admin::EXPRESSION_CACHE_FORCE_COMPACTION_FUEL)
349        .add(&crate::cli::admin::EXPRESSION_CACHE_FORCE_COMPACTION_WAIT)
350        .add(&crate::fetch::FETCH_SEMAPHORE_COST_ADJUSTMENT)
351        .add(&crate::fetch::FETCH_SEMAPHORE_PERMIT_ADJUSTMENT)
352        .add(&crate::fetch::VALIDATE_PART_BOUNDS_ON_READ)
353        .add(&crate::fetch::OPTIMIZE_IGNORED_DATA_FETCH)
354        .add(&crate::internal::cache::BLOB_CACHE_MEM_LIMIT_BYTES)
355        .add(&crate::internal::cache::BLOB_CACHE_SCALE_WITH_THREADS)
356        .add(&crate::internal::cache::BLOB_CACHE_SCALE_FACTOR_BYTES)
357        .add(&crate::internal::compact::COMPACTION_MINIMUM_TIMEOUT)
358        .add(&crate::internal::compact::COMPACTION_CHECK_PROCESS_FLAG)
359        .add(&crate::internal::machine::CLAIM_UNCLAIMED_COMPACTIONS)
360        .add(&crate::internal::machine::CLAIM_COMPACTION_PERCENT)
361        .add(&crate::internal::machine::CLAIM_COMPACTION_MIN_VERSION)
362        .add(&crate::internal::machine::NEXT_LISTEN_BATCH_RETRYER_CLAMP)
363        .add(&crate::internal::machine::NEXT_LISTEN_BATCH_RETRYER_FIXED_SLEEP)
364        .add(&crate::internal::machine::NEXT_LISTEN_BATCH_RETRYER_INITIAL_BACKOFF)
365        .add(&crate::internal::machine::NEXT_LISTEN_BATCH_RETRYER_MULTIPLIER)
366        .add(&crate::internal::state::ROLLUP_THRESHOLD)
367        .add(&crate::internal::state::ROLLUP_USE_ACTIVE_ROLLUP)
368        .add(&crate::internal::state::GC_FALLBACK_THRESHOLD_MS)
369        .add(&crate::internal::state::GC_USE_ACTIVE_GC)
370        .add(&crate::internal::state::GC_MIN_VERSIONS)
371        .add(&crate::internal::state::GC_MAX_VERSIONS)
372        .add(&crate::internal::state::ROLLUP_FALLBACK_THRESHOLD_MS)
373        .add(&crate::internal::state::ENABLE_INCREMENTAL_COMPACTION)
374        .add(&crate::operators::STORAGE_SOURCE_DECODE_FUEL)
375        .add(&crate::read::READER_LEASE_DURATION)
376        .add(&crate::rpc::PUBSUB_CLIENT_ENABLED)
377        .add(&crate::rpc::PUBSUB_PUSH_DIFF_ENABLED)
378        .add(&crate::rpc::PUBSUB_SAME_PROCESS_DELEGATE_ENABLED)
379        .add(&crate::rpc::PUBSUB_CONNECT_ATTEMPT_TIMEOUT)
380        .add(&crate::rpc::PUBSUB_REQUEST_TIMEOUT)
381        .add(&crate::rpc::PUBSUB_CONNECT_MAX_BACKOFF)
382        .add(&crate::rpc::PUBSUB_CLIENT_SENDER_CHANNEL_SIZE)
383        .add(&crate::rpc::PUBSUB_CLIENT_RECEIVER_CHANNEL_SIZE)
384        .add(&crate::rpc::PUBSUB_SERVER_CONNECTION_CHANNEL_SIZE)
385        .add(&crate::rpc::PUBSUB_STATE_CACHE_SHARD_REF_CHANNEL_SIZE)
386        .add(&crate::rpc::PUBSUB_RECONNECT_BACKOFF)
387        .add(&crate::stats::STATS_AUDIT_PERCENT)
388        .add(&crate::stats::STATS_AUDIT_PANIC)
389        .add(&crate::stats::STATS_BUDGET_BYTES)
390        .add(&crate::stats::STATS_COLLECTION_ENABLED)
391        .add(&crate::stats::STATS_FILTER_ENABLED)
392        .add(&crate::stats::STATS_UNTRIMMABLE_COLUMNS_EQUALS)
393        .add(&crate::stats::STATS_UNTRIMMABLE_COLUMNS_PREFIX)
394        .add(&crate::stats::STATS_UNTRIMMABLE_COLUMNS_SUFFIX)
395        .add(&crate::fetch::PART_DECODE_FORMAT)
396        .add(&crate::fetch::PART_DECODE_BATCH_ROWS)
397        .add(&crate::write::COMBINE_INLINE_WRITES)
398        .add(&crate::write::VALIDATE_PART_BOUNDS_ON_WRITE)
399}
400
401impl PersistConfig {
402    pub(crate) const DEFAULT_FALLBACK_ROLLUP_THRESHOLD_MULTIPLIER: usize = 3;
403
404    pub fn set_state_versions_recent_live_diffs_limit(&self, val: usize) {
405        self.set_config(&STATE_VERSIONS_RECENT_LIVE_DIFFS_LIMIT, val);
406    }
407}
408
409/// Sets the maximum size of the connection pool that is used by consensus.
410///
411/// Requires a restart of the process to take effect.
412pub const CONSENSUS_CONNECTION_POOL_MAX_SIZE: Config<usize> = Config::new(
413    "persist_consensus_connection_pool_max_size",
414    50,
415    "The maximum size the connection pool to Postgres/CRDB will grow to.",
416    ParameterScope::Environment,
417);
418
419/// Sets the maximum amount of time we'll wait to acquire a connection from
420/// the connection pool.
421///
422/// Requires a restart of the process to take effect.
423const CONSENSUS_CONNECTION_POOL_MAX_WAIT: Config<Duration> = Config::new(
424    "persist_consensus_connection_pool_max_wait",
425    Duration::from_secs(60),
426    "The amount of time we'll wait for a connection to become available.",
427    ParameterScope::Environment,
428);
429
430/// The minimum TTL of a connection to Postgres/CRDB before it is proactively
431/// terminated. Connections are routinely culled to balance load against the
432/// downstream database.
433const CONSENSUS_CONNECTION_POOL_TTL: Config<Duration> = Config::new(
434    "persist_consensus_connection_pool_ttl",
435    Duration::from_secs(300),
436    "\
437    The minimum TTL of a Consensus connection to Postgres/CRDB before it is \
438    proactively terminated",
439    ParameterScope::Environment,
440);
441
442/// The minimum time between TTLing connections to Postgres/CRDB. This delay is
443/// used to stagger reconnections to avoid stampedes and high tail latencies.
444/// This value should be much less than `consensus_connection_pool_ttl` so that
445/// reconnections are biased towards terminating the oldest connections first. A
446/// value of `consensus_connection_pool_ttl /
447/// consensus_connection_pool_max_size` is likely a good place to start so that
448/// all connections are rotated when the pool is fully used.
449const CONSENSUS_CONNECTION_POOL_TTL_STAGGER: Config<Duration> = Config::new(
450    "persist_consensus_connection_pool_ttl_stagger",
451    Duration::from_secs(6),
452    "The minimum time between TTLing Consensus connections to Postgres/CRDB.",
453    ParameterScope::Environment,
454);
455
456/// The duration to wait for a Consensus Postgres/CRDB connection to be made
457/// before retrying.
458pub const CRDB_CONNECT_TIMEOUT: Config<Duration> = Config::new(
459    "crdb_connect_timeout",
460    Duration::from_secs(5),
461    "The time to connect to CockroachDB before timing out and retrying.",
462    ParameterScope::Environment,
463);
464
465/// The TCP user timeout for a Consensus Postgres/CRDB connection. Specifies the
466/// amount of time that transmitted data may remain unacknowledged before the
467/// TCP connection is forcibly closed.
468pub const CRDB_TCP_USER_TIMEOUT: Config<Duration> = Config::new(
469    "crdb_tcp_user_timeout",
470    Duration::from_secs(30),
471    "\
472    The TCP timeout for connections to CockroachDB. Specifies the amount of \
473    time that transmitted data may remain unacknowledged before the TCP \
474    connection is forcibly closed.",
475    ParameterScope::Environment,
476);
477
478pub const CRDB_KEEPALIVES_IDLE: Config<Duration> = Config::new(
479    "crdb_keepalives_idle",
480    Duration::from_secs(10),
481    "\
482    The amount of idle time before a TCP keepalive packet is sent on CRDB \
483    connections.",
484    ParameterScope::Environment,
485);
486
487pub const CRDB_KEEPALIVES_INTERVAL: Config<Duration> = Config::new(
488    "crdb_keepalives_interval",
489    Duration::from_secs(5),
490    "The time interval between TCP keepalive probes on CRDB connections.",
491    ParameterScope::Environment,
492);
493
494pub const CRDB_KEEPALIVES_RETRIES: Config<u32> = Config::new(
495    "crdb_keepalives_retries",
496    5,
497    "\
498    The maximum number of TCP keepalive probes that will be sent before \
499    dropping a CRDB connection.",
500    ParameterScope::Environment,
501);
502
503/// Migrate the txns code to use the critical since when opening a new read handle.
504pub const USE_CRITICAL_SINCE_TXN: Config<bool> = Config::new(
505    "persist_use_critical_since_txn",
506    true,
507    "Use the critical since (instead of the overall since) when initializing a subscribe.",
508    ParameterScope::Environment,
509);
510
511/// Migrate the catalog to use the critical since when opening a new read handle.
512pub const USE_CRITICAL_SINCE_CATALOG: Config<bool> = Config::new(
513    "persist_use_critical_since_catalog",
514    false,
515    "Use the critical since (instead of the overall since) for the Persist-backed catalog.",
516    ParameterScope::Environment,
517);
518
519/// Migrate the persist source to use the critical since when opening a new read handle.
520pub const USE_CRITICAL_SINCE_SOURCE: Config<bool> = Config::new(
521    "persist_use_critical_since_source",
522    false,
523    "Use the critical since (instead of the overall since) in the Persist source.",
524    ParameterScope::Environment,
525);
526
527/// While the source is catching up to the shard upper observed at hydration,
528/// coalesce per-batch frontier downgrades until at least this many encoded
529/// bytes have been emitted at the held capability. `0` disables coalescing and
530/// restores the per-batch behavior.
531pub const SOURCE_HYDRATION_FRONTIER_COALESCE_BYTES: Config<usize> = Config::new(
532    "persist_source_hydration_frontier_coalesce_bytes",
533    0,
534    "While catching up to the hydration-time upper, the persist source coalesces \
535     frontier downgrades until this many encoded bytes have been emitted (0 disables).",
536    ParameterScope::Environment,
537);
538
539/// Maximum number of part fetches the persist source issues concurrently, per
540/// worker. Concurrent fetches amortize the blob-store round-trip, which
541/// dominates when there are many small parts (e.g. a fine-grained hydration
542/// replay). `1` keeps the previous serial behavior.
543pub const SOURCE_FETCH_CONCURRENCY: Config<usize> = Config::new(
544    "persist_source_fetch_concurrency",
545    1,
546    "Maximum number of part fetches the persist source issues concurrently per worker \
547     (1 = serial).",
548    ParameterScope::Environment,
549);
550
551/// Migrate snapshots to use the critical since when opening a new read handle.
552pub const USE_CRITICAL_SINCE_SNAPSHOT: Config<bool> = Config::new(
553    "persist_use_critical_since_snapshot",
554    false,
555    "Use the critical since (instead of the overall since) when taking snapshots in the controller or in fast-path peeks.",
556    ParameterScope::Environment,
557);
558
559/// The maximum number of parts (s3 blobs) that [crate::batch::BatchBuilder]
560/// will pipeline before back-pressuring [crate::batch::BatchBuilder::add]
561/// calls on previous ones finishing.
562pub const BATCH_BUILDER_MAX_OUTSTANDING_PARTS: Config<usize> = Config::new(
563    "persist_batch_builder_max_outstanding_parts",
564    2,
565    "The number of writes a batch builder can have outstanding before we slow down the writer.",
566    ParameterScope::Environment,
567);
568
569/// In Compactor::compact_and_apply, we do the compaction (don't skip it)
570/// if the number of inputs is at least this many. Compaction is performed
571/// if any of the heuristic criteria are met (they are OR'd).
572pub const COMPACTION_HEURISTIC_MIN_INPUTS: Config<usize> = Config::new(
573    "persist_compaction_heuristic_min_inputs",
574    8,
575    "Don't skip compaction if we have more than this many hollow batches as input.",
576    ParameterScope::Environment,
577);
578
579/// In Compactor::compact_and_apply, we do the compaction (don't skip it)
580/// if the number of batch parts is at least this many. Compaction is performed
581/// if any of the heuristic criteria are met (they are OR'd).
582pub const COMPACTION_HEURISTIC_MIN_PARTS: Config<usize> = Config::new(
583    "persist_compaction_heuristic_min_parts",
584    8,
585    "Don't skip compaction if we have more than this many parts as input.",
586    ParameterScope::Environment,
587);
588
589/// In Compactor::compact_and_apply, we do the compaction (don't skip it)
590/// if the number of updates is at least this many. Compaction is performed
591/// if any of the heuristic criteria are met (they are OR'd).
592pub const COMPACTION_HEURISTIC_MIN_UPDATES: Config<usize> = Config::new(
593    "persist_compaction_heuristic_min_updates",
594    1024,
595    "Don't skip compaction if we have more than this many updates as input.",
596    ParameterScope::Environment,
597);
598
599/// The upper bound on compaction's memory consumption. The value must be at
600/// least 4*`blob_target_size`. Increasing this value beyond the minimum allows
601/// compaction to merge together more runs at once, providing greater
602/// consolidation of updates, at the cost of greater memory usage.
603pub const COMPACTION_MEMORY_BOUND_BYTES: Config<usize> = Config::new(
604    "persist_compaction_memory_bound_bytes",
605    1024 * MiB,
606    "Attempt to limit compaction to this amount of memory.",
607    ParameterScope::Environment,
608);
609
610/// The maximum number of concurrent blob deletes during garbage collection.
611pub const GC_BLOB_DELETE_CONCURRENCY_LIMIT: Config<usize> = Config::new(
612    "persist_gc_blob_delete_concurrency_limit",
613    32,
614    "Limit the number of concurrent deletes GC can perform to this threshold.",
615    ParameterScope::Environment,
616);
617
618/// The # of diffs to initially scan when fetching the latest consensus state, to
619/// determine which requests go down the fast vs slow path. Should be large enough
620/// to fetch all live diffs in the steady-state, and small enough to query Consensus
621/// at high volume. Steady-state usage should accommodate readers that require
622/// seqno-holds for reasonable amounts of time, which to start we say is 10s of minutes.
623///
624/// This value ought to be defined in terms of `NEED_ROLLUP_THRESHOLD` to approximate
625/// when we expect rollups to be written and therefore when old states will be truncated
626/// by GC.
627pub const STATE_VERSIONS_RECENT_LIVE_DIFFS_LIMIT: Config<usize> = Config::new(
628    "persist_state_versions_recent_live_diffs_limit",
629    30 * 128,
630    "Fetch this many diffs when fetching recent diffs.",
631    ParameterScope::Environment,
632);
633
634/// The maximum number of concurrent state fetches during usage computation.
635pub const USAGE_STATE_FETCH_CONCURRENCY_LIMIT: Config<usize> = Config::new(
636    "persist_usage_state_fetch_concurrency_limit",
637    8,
638    "Limit the concurrency in of fetching in the perioding Persist-storage-usage calculation.",
639    ParameterScope::Environment,
640);
641
642impl PostgresClientKnobs for PersistConfig {
643    fn connection_pool_max_size(&self) -> usize {
644        CONSENSUS_CONNECTION_POOL_MAX_SIZE.get(self)
645    }
646
647    fn connection_pool_max_wait(&self) -> Option<Duration> {
648        Some(CONSENSUS_CONNECTION_POOL_MAX_WAIT.get(self))
649    }
650
651    fn connection_pool_ttl(&self) -> Duration {
652        CONSENSUS_CONNECTION_POOL_TTL.get(self)
653    }
654
655    fn connection_pool_ttl_stagger(&self) -> Duration {
656        CONSENSUS_CONNECTION_POOL_TTL_STAGGER.get(self)
657    }
658
659    fn connect_timeout(&self) -> Duration {
660        CRDB_CONNECT_TIMEOUT.get(self)
661    }
662
663    fn tcp_user_timeout(&self) -> Duration {
664        CRDB_TCP_USER_TIMEOUT.get(self)
665    }
666
667    fn keepalives_idle(&self) -> Duration {
668        CRDB_KEEPALIVES_IDLE.get(self)
669    }
670
671    fn keepalives_interval(&self) -> Duration {
672        CRDB_KEEPALIVES_INTERVAL.get(self)
673    }
674
675    fn keepalives_retries(&self) -> u32 {
676        CRDB_KEEPALIVES_RETRIES.get(self)
677    }
678
679    fn statement_timeout(&self) -> Duration {
680        // Persist consensus does not use a server-side statement timeout.
681        Duration::ZERO
682    }
683}
684
685#[derive(Copy, Clone, Debug, Eq, PartialEq, Arbitrary, Serialize, Deserialize)]
686pub struct RetryParameters {
687    pub fixed_sleep: Duration,
688    pub initial_backoff: Duration,
689    pub multiplier: u32,
690    pub clamp: Duration,
691}
692
693impl RetryParameters {
694    pub fn persist_defaults() -> Self {
695        Self {
696            fixed_sleep: Duration::ZERO,
697            // Chosen to meet the following arbitrary criteria: a power of two
698            // that's close to the AWS Aurora latency of 6ms.
699            initial_backoff: Duration::from_millis(4),
700            multiplier: 2,
701            // Chosen to meet the following arbitrary criteria: between 10s and
702            // 60s.
703            clamp: Duration::from_secs(16),
704        }
705    }
706
707    pub(crate) fn into_retry(self, now: SystemTime) -> Retry {
708        let seed = now
709            .duration_since(UNIX_EPOCH)
710            .map_or(0, |x| u64::from(x.subsec_nanos()));
711        Retry {
712            fixed_sleep: self.fixed_sleep,
713            initial_backoff: self.initial_backoff,
714            multiplier: self.multiplier,
715            clamp_backoff: self.clamp,
716            seed,
717        }
718    }
719}
720
721pub(crate) const BLOB_OPERATION_TIMEOUT: Config<Duration> = Config::new(
722    "persist_blob_operation_timeout",
723    Duration::from_secs(180),
724    "Maximum time allowed for a network call, including retry attempts.",
725    ParameterScope::Environment,
726);
727
728pub(crate) const BLOB_OPERATION_ATTEMPT_TIMEOUT: Config<Duration> = Config::new(
729    "persist_blob_operation_attempt_timeout",
730    Duration::from_secs(90),
731    "Maximum time allowed for a single network call.",
732    ParameterScope::Environment,
733);
734
735pub(crate) const BLOB_CONNECT_TIMEOUT: Config<Duration> = Config::new(
736    "persist_blob_connect_timeout",
737    Duration::from_secs(7),
738    "Maximum time to wait for a socket connection to be made.",
739    ParameterScope::Environment,
740);
741
742pub(crate) const BLOB_READ_TIMEOUT: Config<Duration> = Config::new(
743    "persist_blob_read_timeout",
744    Duration::from_secs(10),
745    "Maximum time to wait to read the first byte of a response, including connection time.",
746    ParameterScope::Environment,
747);
748
749/// Which persist shard metrics this process exports, one of the
750/// [`ShardMetricsExport`] values.
751///
752/// Read on every metrics scrape, so a change takes effect on the next scrape.
753/// A scrape reads it twice, once collecting the aggregates and once filtering
754/// the families, so a change between the two mixes the modes for that scrape.
755pub const SHARD_METRICS: Config<&'static str> = Config::new(
756    "persist_shard_metrics",
757    ShardMetricsExport::default().as_str(),
758    "Which persist shard metrics to export: 'per_shard' (the `mz_persist_shard_*` families \
759    with one series per shard), 'summary' (their bounded `_percentile`, `_sum`, `_topk` and \
760    per-version aggregates), 'both', or 'none'. Shard counts and process-level counters are \
761    always exported.",
762    ParameterScope::Environment,
763);
764
765/// The values of [`SHARD_METRICS`].
766#[derive(Debug, Clone, Copy, PartialEq, Eq)]
767pub enum ShardMetricsExport {
768    None,
769    Summary,
770    PerShard,
771    Both,
772}
773
774impl ShardMetricsExport {
775    pub const ALL: [Self; 4] = [Self::None, Self::Summary, Self::PerShard, Self::Both];
776
777    /// The value [`SHARD_METRICS`] defaults to, whose series count is bounded
778    /// by the per-version aggregates rather than the shard count.
779    pub const fn default() -> Self {
780        Self::Summary
781    }
782
783    pub const fn as_str(self) -> &'static str {
784        match self {
785            Self::None => "none",
786            Self::Summary => "summary",
787            Self::PerShard => "per_shard",
788            Self::Both => "both",
789        }
790    }
791
792    /// Parses a value of [`SHARD_METRICS`], or `None` if it is unrecognized.
793    pub fn parse(s: &str) -> Option<Self> {
794        Self::ALL.into_iter().find(|v| v.as_str() == s)
795    }
796
797    /// Parses a value of [`SHARD_METRICS`], falling back to the default if it
798    /// is unrecognized.
799    ///
800    /// Silent: both callers run per scrape, so a typo would log forever.
801    /// `PerShardMetricsFilter` reports it once per change instead.
802    pub fn from_str_lossy(s: &str) -> Self {
803        Self::parse(s).unwrap_or_else(Self::default)
804    }
805
806    pub fn get(configs: &ConfigSet) -> Self {
807        Self::from_str_lossy(&SHARD_METRICS.get(configs))
808    }
809
810    /// Whether to export the per-shard families.
811    pub fn per_shard(self) -> bool {
812        matches!(self, Self::PerShard | Self::Both)
813    }
814
815    /// Whether to export the bounded aggregates of the per-shard families.
816    pub fn summary(self) -> bool {
817        matches!(self, Self::Summary | Self::Both)
818    }
819}
820
821/// Per-shard series to keep exporting while [`SHARD_METRICS`] leaves the
822/// per-shard families out.
823///
824/// A series is kept if its `shard` or `name` label matches this regex. The
825/// match is unanchored, so `s1` also matches `s10`: use `^...$` for an exact
826/// match. Empty keeps none, as does an invalid regex.
827pub const PER_SHARD_METRICS_ENABLE_REGEX: Config<&'static str> = Config::new(
828    "persist_per_shard_metrics_enable_regex",
829    "",
830    "While `persist_shard_metrics` is 'summary' or 'none', keep exporting the per-shard series \
831    whose `shard` or `name` label matches this regex (unanchored). Empty or invalid keeps \
832    none.",
833    ParameterScope::Environment,
834);
835
836/// Whether to export `mz_persist_shard_info`, one series per shard, so that
837/// dashboards can list the shards when [`SHARD_METRICS`] leaves the per-shard
838/// families out.
839pub const SHARD_INFO_METRIC: Config<bool> = Config::new(
840    "persist_shard_info_metric",
841    false,
842    "Whether to export `mz_persist_shard_info`, which lists this process's shards by `shard` \
843    and `name` with one series each.",
844    ParameterScope::Environment,
845);
846
847impl BlobKnobs for PersistConfig {
848    fn operation_timeout(&self) -> Duration {
849        BLOB_OPERATION_TIMEOUT.get(self)
850    }
851
852    fn operation_attempt_timeout(&self) -> Duration {
853        BLOB_OPERATION_ATTEMPT_TIMEOUT.get(self)
854    }
855
856    fn connect_timeout(&self) -> Duration {
857        BLOB_CONNECT_TIMEOUT.get(self)
858    }
859
860    fn read_timeout(&self) -> Duration {
861        BLOB_READ_TIMEOUT.get(self)
862    }
863
864    fn is_cc_active(&self) -> bool {
865        self.is_cc_active
866    }
867
868    fn http_pool(&self) -> BlobHttpPool {
869        BlobHttpPool::get(self)
870    }
871
872    fn http_pool_max_in_flight_per_client(&self) -> usize {
873        BLOB_HTTP_POOL_MAX_IN_FLIGHT_PER_CLIENT.get(self)
874    }
875}
876
877/// If persist gets some encoded ProtoState from the future (e.g. two versions of
878/// code are running simultaneously against the same shard), it might have a
879/// field that the current code doesn't know about. This would be silently
880/// discarded at proto decode time: our Proto library can't handle unknown fields,
881/// and old versions of code might not be able to respect the semantics of the new
882/// fields even if they did.
883///
884/// [1]: https://developers.google.com/protocol-buffers/docs/proto3#unknowns
885///
886/// To detect the bad situation and disallow it, we tag every version of state
887/// written to consensus with the version of code it's compatible with. Then at
888/// decode time, we're able to compare the current version against any we receive
889/// and assert as necessary. The current version is typically the version of code
890/// used to write the state, but it may be lower when code is intentionally emulating
891/// an older version during eg. a graceful upgrade process.
892///
893/// We could do the same for blob data, but it shouldn't be necessary. Any blob
894/// data we read is going to be because we fetched it using a pointer stored in
895/// some persist state. If we can handle the state, we can handle the blobs it
896/// references, too.
897pub fn code_can_read_data(code_version: &Version, data_version: &Version) -> bool {
898    // For now, Persist can read arbitrarily old state data.
899    // We expect to add a floor to this in future versions.
900    code_version.cmp_precedence(data_version).is_ge()
901}
902
903/// Can the given version of the code generate data that older versions can understand?
904/// Imagine the case of eg. garbage collection after a version upgrade... we may need to read old
905/// diffs to be able to find blobs to delete, even if we no longer have code to generate data in
906/// that format.
907pub fn code_can_write_data(code_version: &Version, data_version: &Version) -> bool {
908    if !code_can_read_data(code_version, data_version) {
909        return false;
910    }
911
912    if code_version.major == 0 && code_version.minor <= SELF_MANAGED_VERSIONS[1].minor {
913        // This code was added well after the last ad-hoc version was released,
914        // so we don't strictly model compatibility with earlier releases.
915        true
916    } else if code_version.major == 0 {
917        // Self-managed versions 25.2+ must be upgradeable from 25.1+.
918        SELF_MANAGED_VERSIONS[0]
919            .cmp_precedence(data_version)
920            .is_le()
921    } else if code_version.major <= 26 {
922        // Versions 26.x must be upgradeable from the last pre-1.0 release.
923        SELF_MANAGED_VERSIONS[1]
924            .cmp_precedence(data_version)
925            .is_le()
926    } else {
927        // Otherwise, the data must be from at earliest the _previous_ major version.
928        code_version.major - 1 <= data_version.major
929    }
930}
mz_persist_client/cfg.rs

mz_persist_client/
cfg.rs
