1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
// Copyright Materialize, Inc. and contributors. All rights reserved.
//
// Use of this software is governed by the Business Source License
// included in the LICENSE file.
//
// As of the Change Date specified in that file, in accordance with
// the Business Source License, use of this software will be governed
// by the Apache License, Version 2.0.

//! Management of arrangements across dataflows.

use std::any::Any;
use std::collections::HashMap;
use std::rc::Rc;

use differential_dataflow::operators::arrange::TraceAgent;
use differential_dataflow::trace::implementations::ord::{OrdKeySpine, OrdValSpine};
use differential_dataflow::trace::TraceReader;
use ore::metric;
use ore::metrics::{
    CounterVec, CounterVecExt, DeleteOnDropCounter, DeleteOnDropGauge, GaugeVecExt,
    MetricsRegistry, UIntGaugeVec,
};
use timely::progress::frontier::{Antichain, AntichainRef};

use dataflow_types::DataflowError;
use expr::GlobalId;
use repr::{Diff, Row, Timestamp};

pub type RowSpine<K, V, T, R, O = usize> = OrdValSpine<K, V, T, R, O>;
pub type ErrSpine<K, T, R, O = usize> = OrdKeySpine<K, T, R, O>;

pub type TraceRowHandle<K, V, T, R> = TraceAgent<RowSpine<K, V, T, R>>;
pub type TraceErrHandle<K, T, R> = TraceAgent<ErrSpine<K, T, R>>;
pub type KeysValsHandle = TraceRowHandle<Row, Row, Timestamp, Diff>;
pub type ErrsHandle = TraceErrHandle<DataflowError, Timestamp, Diff>;

use prometheus::core::{AtomicF64, AtomicU64};
use std::time::Instant;

/// Base metrics for arrangements.
#[derive(Clone, Debug)]
pub struct TraceMetrics {
    total_maintenance_time: CounterVec,
    doing_maintenance: UIntGaugeVec,
}

impl TraceMetrics {
    pub(crate) fn register_with(registry: &MetricsRegistry) -> Self {
        Self {
            total_maintenance_time: registry.register(metric!(
                name: "mz_arrangement_maintenance_seconds_total",
                help: "The total time spent maintaining an arrangement",
                var_labels: ["worker_id", "arrangement_id"],
            )),
            doing_maintenance: registry.register(metric!(
                name: "mz_arrangement_maintenance_active_info",
                help: "Whether or not maintenance is currently occurring",
                var_labels: ["worker_id"],
            )),
        }
    }

    fn maintenance_time_metric(
        &self,
        worker_id: usize,
        id: GlobalId,
    ) -> DeleteOnDropCounter<'static, AtomicF64, Vec<String>> {
        self.total_maintenance_time
            .get_delete_on_drop_counter(vec![worker_id.to_string(), id.to_string()])
    }

    fn maintenance_flag_metric(
        &self,
        worker_id: usize,
    ) -> DeleteOnDropGauge<'static, AtomicU64, Vec<String>> {
        self.doing_maintenance
            .get_delete_on_drop_gauge(vec![worker_id.to_string()])
    }
}

struct MaintenanceMetrics {
    /// total time spent doing maintenance. More useful in the general case.
    total_maintenance_time: DeleteOnDropCounter<'static, AtomicF64, Vec<String>>,
}

impl MaintenanceMetrics {
    fn new(metrics: &TraceMetrics, worker_id: usize, arrangement_id: GlobalId) -> Self {
        MaintenanceMetrics {
            total_maintenance_time: metrics.maintenance_time_metric(worker_id, arrangement_id),
        }
    }
}

/// A `TraceManager` stores maps from global identifiers to the primary arranged
/// representation of that collection.
pub struct TraceManager {
    pub traces: HashMap<GlobalId, TraceBundle>,
    worker_id: usize,
    maintenance_metrics: HashMap<GlobalId, MaintenanceMetrics>,
    /// 1 if this worker is currently doing maintenance.
    ///
    /// If maintenance turns out to take a very long time, this will allow us
    /// to gain a sense that materialize is stuck on maintenance before the
    /// maintenance completes
    doing_maintenance: DeleteOnDropGauge<'static, AtomicU64, Vec<String>>,
    metrics: TraceMetrics,
}

impl TraceManager {
    pub fn new(metrics: TraceMetrics, worker_id: usize) -> Self {
        let doing_maintenance = metrics.maintenance_flag_metric(worker_id);
        TraceManager {
            traces: HashMap::new(),
            worker_id,
            metrics,
            maintenance_metrics: HashMap::new(),
            doing_maintenance,
        }
    }

    /// performs maintenance work on the managed traces.
    ///
    /// In particular, this method enables the physical merging of batches, so that at most a logarithmic
    /// number of batches need to be maintained. Any new batches introduced after this method is called
    /// will not be physically merged until the method is called again. This is mostly due to limitations
    /// of differential dataflow, which requires users to perform this explicitly; if that changes we may
    /// be able to remove this code.
    pub fn maintenance(&mut self) {
        let mut antichain = Antichain::new();
        for (arrangement_id, bundle) in self.traces.iter_mut() {
            // Update maintenance metrics
            // Entry is guaranteed to exist as it gets created when we initialize the partition.
            let maintenance_metrics = self.maintenance_metrics.get_mut(arrangement_id).unwrap();

            // signal that maintenance is happening
            self.doing_maintenance.set(1);
            let now = Instant::now();

            bundle.oks.read_upper(&mut antichain);
            bundle.oks.set_physical_compaction(antichain.borrow());
            bundle.errs.read_upper(&mut antichain);
            bundle.errs.set_physical_compaction(antichain.borrow());

            maintenance_metrics
                .total_maintenance_time
                .inc_by(now.elapsed().as_secs_f64());
            // signal that maintenance has ended
            self.doing_maintenance.set(0);
        }
    }

    /// Enables compaction of traces associated with the identifier.
    ///
    /// Compaction may not occur immediately, but once this method is called the
    /// associated traces may not accumulate to the correct quantities for times
    /// not in advance of `frontier`. Users should take care to only rely on
    /// accumulations at times in advance of `frontier`.
    pub fn allow_compaction(&mut self, id: GlobalId, frontier: AntichainRef<Timestamp>) {
        if let Some(bundle) = self.traces.get_mut(&id) {
            bundle.oks.set_logical_compaction(frontier);
            bundle.errs.set_logical_compaction(frontier);
        }
    }

    /// Returns a reference to the trace for `id`, should it exist.
    pub fn get(&self, id: &GlobalId) -> Option<&TraceBundle> {
        self.traces.get(&id)
    }

    /// Returns a mutable reference to the trace for `id`, should it
    /// exist.
    pub fn get_mut(&mut self, id: &GlobalId) -> Option<&mut TraceBundle> {
        self.traces.get_mut(&id)
    }

    /// Binds the arrangement for `id` to `trace`.
    pub fn set(&mut self, id: GlobalId, trace: TraceBundle) {
        self.maintenance_metrics.insert(
            id,
            MaintenanceMetrics::new(&self.metrics, self.worker_id, id),
        );
        self.traces.insert(id, trace);
    }

    /// Removes the trace for `id`.
    pub fn del_trace(&mut self, id: &GlobalId) -> bool {
        self.maintenance_metrics.remove(id);
        self.traces.remove(&id).is_some()
    }

    /// Removes all managed traces.
    pub fn del_all_traces(&mut self) {
        self.maintenance_metrics.clear();
        self.traces.clear();
    }
}

/// Bundles together traces for the successful computations (`oks`), the
/// failed computations (`errs`), additional tokens that should share
/// the lifetime of the bundled traces (`to_drop`), and a permutation
/// describing how to reconstruct the original row (`permutation`).
#[derive(Clone)]
pub struct TraceBundle {
    oks: KeysValsHandle,
    errs: ErrsHandle,
    to_drop: Option<Rc<dyn Any>>,
}

impl TraceBundle {
    /// Constructs a new trace bundle out of an `oks` trace and `errs` trace.
    pub fn new(oks: KeysValsHandle, errs: ErrsHandle) -> TraceBundle {
        TraceBundle {
            oks,
            errs,
            to_drop: None,
        }
    }

    /// Adds tokens to be dropped when the trace bundle is dropped.
    pub fn with_drop<T>(self, to_drop: T) -> TraceBundle
    where
        T: 'static,
    {
        TraceBundle {
            to_drop: Some(Rc::new(Box::new(to_drop))),
            ..self
        }
    }

    /// Returns a mutable reference to the `oks` trace.
    pub fn oks_mut(&mut self) -> &mut KeysValsHandle {
        &mut self.oks
    }

    /// Returns a mutable reference to the `errs` trace.
    pub fn errs_mut(&mut self) -> &mut ErrsHandle {
        &mut self.errs
    }

    /// Returns a reference to the `to_drop` tokens.
    pub fn to_drop(&self) -> &Option<Rc<dyn Any>> {
        &self.to_drop
    }
}