2020-03-20 00:35:42 -04:00
|
|
|
// This file is part of Moonfire NVR, a security camera network video recorder.
|
2021-02-17 16:28:48 -05:00
|
|
|
// Copyright (C) 2020 The Moonfire NVR Authors; see AUTHORS and LICENSE.txt.
|
|
|
|
// SPDX-License-Identifier: GPL-v3.0-or-later WITH GPL-3.0-linking-exception.
|
2020-03-20 00:35:42 -04:00
|
|
|
|
|
|
|
/// Upgrades a version 4 schema to a version 5 schema.
|
|
|
|
use byteorder::{BigEndian, ByteOrder, WriteBytesExt};
|
2021-02-17 01:15:54 -05:00
|
|
|
use failure::{bail, format_err, Error, ResultExt};
|
2020-03-20 00:35:42 -04:00
|
|
|
use h264_reader::avcc::AvcDecoderConfigurationRecord;
|
|
|
|
use rusqlite::{named_params, params};
|
|
|
|
use std::convert::{TryFrom, TryInto};
|
|
|
|
|
|
|
|
// Copied from src/h264.rs. h264 stuff really doesn't belong in the db crate, but we do what we
|
|
|
|
// must for schema upgrades.
|
|
|
|
const PIXEL_ASPECT_RATIOS: [((u16, u16), (u16, u16)); 4] = [
|
2021-02-17 01:15:54 -05:00
|
|
|
((320, 240), (4, 3)),
|
2020-03-20 00:35:42 -04:00
|
|
|
((352, 240), (40, 33)),
|
2021-02-17 01:15:54 -05:00
|
|
|
((640, 480), (4, 3)),
|
2020-03-20 00:35:42 -04:00
|
|
|
((704, 480), (40, 33)),
|
|
|
|
];
|
|
|
|
fn default_pixel_aspect_ratio(width: u16, height: u16) -> (u16, u16) {
|
|
|
|
let dims = (width, height);
|
|
|
|
for r in &PIXEL_ASPECT_RATIOS {
|
|
|
|
if r.0 == dims {
|
|
|
|
return r.1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
(1, 1)
|
|
|
|
}
|
|
|
|
|
|
|
|
fn parse(data: &[u8]) -> Result<AvcDecoderConfigurationRecord, Error> {
|
|
|
|
if data.len() < 94 || &data[4..8] != b"avc1" || &data[90..94] != b"avcC" {
|
|
|
|
bail!("data of len {} doesn't have an avcC", data.len());
|
|
|
|
}
|
|
|
|
let avcc_len = BigEndian::read_u32(&data[86..90]);
|
2021-02-17 01:15:54 -05:00
|
|
|
if avcc_len < 8 {
|
|
|
|
// length and type.
|
2020-03-20 00:35:42 -04:00
|
|
|
bail!("invalid avcc len {}", avcc_len);
|
|
|
|
}
|
|
|
|
let end_pos = 86 + usize::try_from(avcc_len)?;
|
|
|
|
if end_pos != data.len() {
|
2021-02-17 01:15:54 -05:00
|
|
|
bail!(
|
|
|
|
"expected avcC to be end of extradata; there are {} more bytes.",
|
|
|
|
data.len() - end_pos
|
|
|
|
);
|
2020-03-20 00:35:42 -04:00
|
|
|
}
|
|
|
|
AvcDecoderConfigurationRecord::try_from(&data[94..end_pos])
|
|
|
|
.map_err(|e| format_err!("Bad AvcDecoderConfigurationRecord: {:?}", e))
|
|
|
|
}
|
|
|
|
|
|
|
|
pub fn run(_args: &super::Args, tx: &rusqlite::Transaction) -> Result<(), Error> {
|
|
|
|
// These create statements match the schema.sql when version 5 was the latest.
|
2021-02-17 01:15:54 -05:00
|
|
|
tx.execute_batch(
|
|
|
|
r#"
|
2020-03-20 23:52:30 -04:00
|
|
|
alter table video_sample_entry rename to old_video_sample_entry;
|
|
|
|
|
|
|
|
create table video_sample_entry (
|
|
|
|
id integer primary key,
|
|
|
|
width integer not null check (width > 0),
|
|
|
|
height integer not null check (height > 0),
|
|
|
|
rfc6381_codec text not null,
|
|
|
|
data blob not null check (length(data) > 86),
|
|
|
|
pasp_h_spacing integer not null default 1 check (pasp_h_spacing > 0),
|
|
|
|
pasp_v_spacing integer not null default 1 check (pasp_v_spacing > 0)
|
|
|
|
);
|
2021-02-17 01:15:54 -05:00
|
|
|
"#,
|
|
|
|
)?;
|
2020-03-20 00:35:42 -04:00
|
|
|
|
2021-02-17 01:15:54 -05:00
|
|
|
let mut insert = tx.prepare(
|
|
|
|
r#"
|
2020-03-20 23:52:30 -04:00
|
|
|
insert into video_sample_entry (id, width, height, rfc6381_codec, data,
|
|
|
|
pasp_h_spacing, pasp_v_spacing)
|
|
|
|
values (:id, :width, :height, :rfc6381_codec, :data,
|
|
|
|
:pasp_h_spacing, :pasp_v_spacing)
|
2021-02-17 01:15:54 -05:00
|
|
|
"#,
|
|
|
|
)?;
|
2020-06-10 22:35:52 -04:00
|
|
|
|
|
|
|
// Only insert still-referenced video sample entries. I've had problems with
|
|
|
|
// no-longer-referenced ones (perhaps from some ancient, buggy version of Moonfire NVR) for
|
|
|
|
// which avcc.create_context(()) fails.
|
2021-02-17 01:15:54 -05:00
|
|
|
let mut stmt = tx.prepare(
|
|
|
|
r#"
|
2020-03-20 00:35:42 -04:00
|
|
|
select
|
|
|
|
id,
|
|
|
|
width,
|
|
|
|
height,
|
2020-03-20 23:52:30 -04:00
|
|
|
rfc6381_codec,
|
2020-03-20 00:35:42 -04:00
|
|
|
data
|
|
|
|
from
|
2020-06-10 22:35:52 -04:00
|
|
|
old_video_sample_entry v
|
|
|
|
where
|
|
|
|
exists (
|
|
|
|
select
|
|
|
|
'x'
|
|
|
|
from
|
|
|
|
recording r
|
|
|
|
where
|
|
|
|
r.video_sample_entry_id = v.id)
|
2021-02-17 01:15:54 -05:00
|
|
|
"#,
|
|
|
|
)?;
|
2020-03-20 00:35:42 -04:00
|
|
|
let mut rows = stmt.query(params![])?;
|
|
|
|
while let Some(row) = rows.next()? {
|
|
|
|
let id: i32 = row.get(0)?;
|
|
|
|
let width: u16 = row.get::<_, i32>(1)?.try_into()?;
|
|
|
|
let height: u16 = row.get::<_, i32>(2)?.try_into()?;
|
2020-03-20 23:52:30 -04:00
|
|
|
let rfc6381_codec: &str = row.get_raw_checked(3)?.as_str()?;
|
|
|
|
let mut data: Vec<u8> = row.get(4)?;
|
2020-03-20 00:35:42 -04:00
|
|
|
let avcc = parse(&data)?;
|
|
|
|
if avcc.num_of_sequence_parameter_sets() != 1 {
|
|
|
|
bail!("Multiple SPSs!");
|
|
|
|
}
|
2021-02-17 01:15:54 -05:00
|
|
|
let ctx = avcc.create_context(()).map_err(|e| {
|
|
|
|
format_err!(
|
|
|
|
"Can't load SPS+PPS for video_sample_entry_id {}: {:?}",
|
|
|
|
id,
|
|
|
|
e
|
|
|
|
)
|
|
|
|
})?;
|
|
|
|
let sps = ctx
|
|
|
|
.sps_by_id(h264_reader::nal::pps::ParamSetId::from_u32(0).unwrap())
|
2020-06-10 22:35:52 -04:00
|
|
|
.ok_or_else(|| format_err!("No SPS 0 for video_sample_entry_id {}", id))?;
|
2021-02-17 01:15:54 -05:00
|
|
|
let pasp = sps
|
|
|
|
.vui_parameters
|
|
|
|
.as_ref()
|
|
|
|
.and_then(|v| v.aspect_ratio_info.as_ref())
|
|
|
|
.and_then(|a| a.clone().get())
|
|
|
|
.unwrap_or_else(|| default_pixel_aspect_ratio(width, height));
|
2020-03-20 00:35:42 -04:00
|
|
|
if pasp != (1, 1) {
|
2021-02-17 01:15:54 -05:00
|
|
|
data.extend_from_slice(b"\x00\x00\x00\x10pasp"); // length + box name
|
2020-03-20 00:35:42 -04:00
|
|
|
data.write_u32::<BigEndian>(pasp.0.into())?;
|
|
|
|
data.write_u32::<BigEndian>(pasp.1.into())?;
|
|
|
|
let len = data.len();
|
|
|
|
BigEndian::write_u32(&mut data[0..4], u32::try_from(len)?);
|
|
|
|
}
|
|
|
|
|
2021-02-17 01:15:54 -05:00
|
|
|
insert.execute_named(named_params! {
|
2020-03-20 00:35:42 -04:00
|
|
|
":id": id,
|
2020-03-20 23:52:30 -04:00
|
|
|
":width": width,
|
|
|
|
":height": height,
|
|
|
|
":rfc6381_codec": rfc6381_codec,
|
2020-03-20 00:35:42 -04:00
|
|
|
":data": &data,
|
|
|
|
":pasp_h_spacing": pasp.0,
|
|
|
|
":pasp_v_spacing": pasp.1,
|
|
|
|
})?;
|
|
|
|
}
|
2021-02-17 01:15:54 -05:00
|
|
|
tx.execute_batch(
|
|
|
|
r#"
|
track cumulative duration and runs
This is useful for a combo scrub bar-based UI (#32) + live view UI (#59)
in a non-obvious way. When constructing a HTML Media Source Extensions
API SourceBuffer, the caller can specify a "mode" of either "segments"
or "sequence":
In "sequence" mode, playback assumes segments are added sequentially.
This is good enough for a live view-only UI (#59) but not for a scrub
bar UI in which you may want to seek backward to a segment you've never
seen before. You will then need to insert a segment out-of-sequence.
Imagine what happens when the user goes forward again until the end of
the segment inserted immediately before it. The user should see the
chronologically next segment or a pause for loading if it's unavailable.
The best approximation of this is to track the mapping of timestamps to
segments and insert a VTTCue with an enter/exit handler that seeks to
the right position. But seeking isn't instantaneous; the user will
likely briefly see first the segment they seeked to before. That's
janky. Additionally, the "canplaythrough" event will behave strangely.
In "segments" mode, playback respects the timestamps we set:
* The obvious choice is to use wall clock timestamps. This is fine if
they're known to be fixed and correct. They're not. The
currently-recording segment may be "unanchored", meaning its start
timestamp is not yet fixed. Older timestamps may overlap if the system
clock was stepped between runs. The latter isn't /too/ bad from a user
perspective, though it's confusing as a developer. We probably will
only end up showing the more recent recording for a given
timestamp anyway. But the former is quite annoying. It means we have
to throw away part of the SourceBuffer that we may want to seek back
(causing UI pauses when that happens) or keep our own spare copy of it
(memory bloat). I'd like to avoid the whole mess.
* Another approach is to use timestamps that are guaranteed to be in
the correct order but that may have gaps. In particular, a timestamp
of (recording_id * max_recording_duration) + time_within_recording.
But again seeking isn't instantaneous. In my experiments, there's a
visible pause between segments that drives me nuts.
* Finally, the approach that led me to this schema change. Use
timestamps that place each segment after the one before, possibly with
an intentional gap between runs (to force a wait where we have an
actual gap). This should make the browser's natural playback behavior
work properly: it never goes to an incorrect place, and it only waits
when/if we want it to. We have to maintain a mapping between its
timestamps and segment ids but that's doable.
This commit is only the schema change; the new data aren't exposed in
the API yet, much less used by a UI.
Note that stream.next_recording_id became stream.cum_recordings. I made
a slight definition change in the process: recording ids for new streams
start at 0 rather than 1. Various tests changed accordingly.
The upgrade process makes a best effort to backfill these new fields,
but of course it doesn't know the total duration or number of runs of
previously deleted rows. That's good enough.
2020-06-09 19:17:32 -04:00
|
|
|
alter table stream rename to old_stream;
|
|
|
|
create table stream (
|
|
|
|
id integer primary key,
|
|
|
|
camera_id integer not null references camera (id),
|
|
|
|
sample_file_dir_id integer references sample_file_dir (id),
|
|
|
|
type text not null check (type in ('main', 'sub')),
|
|
|
|
record integer not null check (record in (1, 0)),
|
|
|
|
rtsp_url text not null,
|
|
|
|
retain_bytes integer not null check (retain_bytes >= 0),
|
|
|
|
flush_if_sec integer not null,
|
|
|
|
cum_recordings integer not null check (cum_recordings >= 0),
|
2020-08-05 00:44:01 -04:00
|
|
|
cum_media_duration_90k integer not null check (cum_media_duration_90k >= 0),
|
track cumulative duration and runs
This is useful for a combo scrub bar-based UI (#32) + live view UI (#59)
in a non-obvious way. When constructing a HTML Media Source Extensions
API SourceBuffer, the caller can specify a "mode" of either "segments"
or "sequence":
In "sequence" mode, playback assumes segments are added sequentially.
This is good enough for a live view-only UI (#59) but not for a scrub
bar UI in which you may want to seek backward to a segment you've never
seen before. You will then need to insert a segment out-of-sequence.
Imagine what happens when the user goes forward again until the end of
the segment inserted immediately before it. The user should see the
chronologically next segment or a pause for loading if it's unavailable.
The best approximation of this is to track the mapping of timestamps to
segments and insert a VTTCue with an enter/exit handler that seeks to
the right position. But seeking isn't instantaneous; the user will
likely briefly see first the segment they seeked to before. That's
janky. Additionally, the "canplaythrough" event will behave strangely.
In "segments" mode, playback respects the timestamps we set:
* The obvious choice is to use wall clock timestamps. This is fine if
they're known to be fixed and correct. They're not. The
currently-recording segment may be "unanchored", meaning its start
timestamp is not yet fixed. Older timestamps may overlap if the system
clock was stepped between runs. The latter isn't /too/ bad from a user
perspective, though it's confusing as a developer. We probably will
only end up showing the more recent recording for a given
timestamp anyway. But the former is quite annoying. It means we have
to throw away part of the SourceBuffer that we may want to seek back
(causing UI pauses when that happens) or keep our own spare copy of it
(memory bloat). I'd like to avoid the whole mess.
* Another approach is to use timestamps that are guaranteed to be in
the correct order but that may have gaps. In particular, a timestamp
of (recording_id * max_recording_duration) + time_within_recording.
But again seeking isn't instantaneous. In my experiments, there's a
visible pause between segments that drives me nuts.
* Finally, the approach that led me to this schema change. Use
timestamps that place each segment after the one before, possibly with
an intentional gap between runs (to force a wait where we have an
actual gap). This should make the browser's natural playback behavior
work properly: it never goes to an incorrect place, and it only waits
when/if we want it to. We have to maintain a mapping between its
timestamps and segment ids but that's doable.
This commit is only the schema change; the new data aren't exposed in
the API yet, much less used by a UI.
Note that stream.next_recording_id became stream.cum_recordings. I made
a slight definition change in the process: recording ids for new streams
start at 0 rather than 1. Various tests changed accordingly.
The upgrade process makes a best effort to backfill these new fields,
but of course it doesn't know the total duration or number of runs of
previously deleted rows. That's good enough.
2020-06-09 19:17:32 -04:00
|
|
|
cum_runs integer not null check (cum_runs >= 0),
|
|
|
|
unique (camera_id, type)
|
|
|
|
);
|
|
|
|
insert into stream
|
|
|
|
select
|
|
|
|
s.id,
|
|
|
|
s.camera_id,
|
|
|
|
s.sample_file_dir_id,
|
|
|
|
s.type,
|
|
|
|
s.record,
|
|
|
|
s.rtsp_url,
|
|
|
|
s.retain_bytes,
|
|
|
|
s.flush_if_sec,
|
|
|
|
s.next_recording_id as cum_recordings,
|
|
|
|
coalesce(sum(r.duration_90k), 0) as cum_duration_90k,
|
|
|
|
coalesce(sum(case when r.run_offset = 0 then 1 else 0 end), 0) as cum_runs
|
|
|
|
from
|
|
|
|
old_stream s left join recording r on (s.id = r.stream_id)
|
|
|
|
group by 1;
|
|
|
|
|
2020-03-20 23:52:30 -04:00
|
|
|
alter table recording rename to old_recording;
|
|
|
|
create table recording (
|
|
|
|
composite_id integer primary key,
|
|
|
|
open_id integer not null,
|
|
|
|
stream_id integer not null references stream (id),
|
|
|
|
run_offset integer not null,
|
|
|
|
flags integer not null,
|
|
|
|
sample_file_bytes integer not null check (sample_file_bytes > 0),
|
|
|
|
start_time_90k integer not null check (start_time_90k > 0),
|
2020-08-05 00:44:01 -04:00
|
|
|
prev_media_duration_90k integer not null check (prev_media_duration_90k >= 0),
|
track cumulative duration and runs
This is useful for a combo scrub bar-based UI (#32) + live view UI (#59)
in a non-obvious way. When constructing a HTML Media Source Extensions
API SourceBuffer, the caller can specify a "mode" of either "segments"
or "sequence":
In "sequence" mode, playback assumes segments are added sequentially.
This is good enough for a live view-only UI (#59) but not for a scrub
bar UI in which you may want to seek backward to a segment you've never
seen before. You will then need to insert a segment out-of-sequence.
Imagine what happens when the user goes forward again until the end of
the segment inserted immediately before it. The user should see the
chronologically next segment or a pause for loading if it's unavailable.
The best approximation of this is to track the mapping of timestamps to
segments and insert a VTTCue with an enter/exit handler that seeks to
the right position. But seeking isn't instantaneous; the user will
likely briefly see first the segment they seeked to before. That's
janky. Additionally, the "canplaythrough" event will behave strangely.
In "segments" mode, playback respects the timestamps we set:
* The obvious choice is to use wall clock timestamps. This is fine if
they're known to be fixed and correct. They're not. The
currently-recording segment may be "unanchored", meaning its start
timestamp is not yet fixed. Older timestamps may overlap if the system
clock was stepped between runs. The latter isn't /too/ bad from a user
perspective, though it's confusing as a developer. We probably will
only end up showing the more recent recording for a given
timestamp anyway. But the former is quite annoying. It means we have
to throw away part of the SourceBuffer that we may want to seek back
(causing UI pauses when that happens) or keep our own spare copy of it
(memory bloat). I'd like to avoid the whole mess.
* Another approach is to use timestamps that are guaranteed to be in
the correct order but that may have gaps. In particular, a timestamp
of (recording_id * max_recording_duration) + time_within_recording.
But again seeking isn't instantaneous. In my experiments, there's a
visible pause between segments that drives me nuts.
* Finally, the approach that led me to this schema change. Use
timestamps that place each segment after the one before, possibly with
an intentional gap between runs (to force a wait where we have an
actual gap). This should make the browser's natural playback behavior
work properly: it never goes to an incorrect place, and it only waits
when/if we want it to. We have to maintain a mapping between its
timestamps and segment ids but that's doable.
This commit is only the schema change; the new data aren't exposed in
the API yet, much less used by a UI.
Note that stream.next_recording_id became stream.cum_recordings. I made
a slight definition change in the process: recording ids for new streams
start at 0 rather than 1. Various tests changed accordingly.
The upgrade process makes a best effort to backfill these new fields,
but of course it doesn't know the total duration or number of runs of
previously deleted rows. That's good enough.
2020-06-09 19:17:32 -04:00
|
|
|
prev_runs integer not null check (prev_runs >= 0),
|
2020-08-05 00:44:01 -04:00
|
|
|
wall_duration_90k integer not null
|
|
|
|
check (wall_duration_90k >= 0 and wall_duration_90k < 5*60*90000),
|
|
|
|
media_duration_delta_90k integer not null,
|
2020-03-20 23:52:30 -04:00
|
|
|
video_samples integer not null check (video_samples > 0),
|
|
|
|
video_sync_samples integer not null check (video_sync_samples > 0),
|
|
|
|
video_sample_entry_id integer references video_sample_entry (id),
|
|
|
|
check (composite_id >> 32 = stream_id)
|
|
|
|
);
|
2021-02-17 01:15:54 -05:00
|
|
|
"#,
|
|
|
|
)?;
|
track cumulative duration and runs
This is useful for a combo scrub bar-based UI (#32) + live view UI (#59)
in a non-obvious way. When constructing a HTML Media Source Extensions
API SourceBuffer, the caller can specify a "mode" of either "segments"
or "sequence":
In "sequence" mode, playback assumes segments are added sequentially.
This is good enough for a live view-only UI (#59) but not for a scrub
bar UI in which you may want to seek backward to a segment you've never
seen before. You will then need to insert a segment out-of-sequence.
Imagine what happens when the user goes forward again until the end of
the segment inserted immediately before it. The user should see the
chronologically next segment or a pause for loading if it's unavailable.
The best approximation of this is to track the mapping of timestamps to
segments and insert a VTTCue with an enter/exit handler that seeks to
the right position. But seeking isn't instantaneous; the user will
likely briefly see first the segment they seeked to before. That's
janky. Additionally, the "canplaythrough" event will behave strangely.
In "segments" mode, playback respects the timestamps we set:
* The obvious choice is to use wall clock timestamps. This is fine if
they're known to be fixed and correct. They're not. The
currently-recording segment may be "unanchored", meaning its start
timestamp is not yet fixed. Older timestamps may overlap if the system
clock was stepped between runs. The latter isn't /too/ bad from a user
perspective, though it's confusing as a developer. We probably will
only end up showing the more recent recording for a given
timestamp anyway. But the former is quite annoying. It means we have
to throw away part of the SourceBuffer that we may want to seek back
(causing UI pauses when that happens) or keep our own spare copy of it
(memory bloat). I'd like to avoid the whole mess.
* Another approach is to use timestamps that are guaranteed to be in
the correct order but that may have gaps. In particular, a timestamp
of (recording_id * max_recording_duration) + time_within_recording.
But again seeking isn't instantaneous. In my experiments, there's a
visible pause between segments that drives me nuts.
* Finally, the approach that led me to this schema change. Use
timestamps that place each segment after the one before, possibly with
an intentional gap between runs (to force a wait where we have an
actual gap). This should make the browser's natural playback behavior
work properly: it never goes to an incorrect place, and it only waits
when/if we want it to. We have to maintain a mapping between its
timestamps and segment ids but that's doable.
This commit is only the schema change; the new data aren't exposed in
the API yet, much less used by a UI.
Note that stream.next_recording_id became stream.cum_recordings. I made
a slight definition change in the process: recording ids for new streams
start at 0 rather than 1. Various tests changed accordingly.
The upgrade process makes a best effort to backfill these new fields,
but of course it doesn't know the total duration or number of runs of
previously deleted rows. That's good enough.
2020-06-09 19:17:32 -04:00
|
|
|
|
|
|
|
// SQLite added window functions in 3.25.0. macOS still ships SQLite 3.24.0 (no support).
|
|
|
|
// Compute cumulative columns by hand.
|
|
|
|
let mut cur_stream_id = None;
|
|
|
|
let mut cum_duration_90k = 0;
|
|
|
|
let mut cum_runs = 0;
|
2021-02-17 01:15:54 -05:00
|
|
|
let mut stmt = tx.prepare(
|
|
|
|
r#"
|
track cumulative duration and runs
This is useful for a combo scrub bar-based UI (#32) + live view UI (#59)
in a non-obvious way. When constructing a HTML Media Source Extensions
API SourceBuffer, the caller can specify a "mode" of either "segments"
or "sequence":
In "sequence" mode, playback assumes segments are added sequentially.
This is good enough for a live view-only UI (#59) but not for a scrub
bar UI in which you may want to seek backward to a segment you've never
seen before. You will then need to insert a segment out-of-sequence.
Imagine what happens when the user goes forward again until the end of
the segment inserted immediately before it. The user should see the
chronologically next segment or a pause for loading if it's unavailable.
The best approximation of this is to track the mapping of timestamps to
segments and insert a VTTCue with an enter/exit handler that seeks to
the right position. But seeking isn't instantaneous; the user will
likely briefly see first the segment they seeked to before. That's
janky. Additionally, the "canplaythrough" event will behave strangely.
In "segments" mode, playback respects the timestamps we set:
* The obvious choice is to use wall clock timestamps. This is fine if
they're known to be fixed and correct. They're not. The
currently-recording segment may be "unanchored", meaning its start
timestamp is not yet fixed. Older timestamps may overlap if the system
clock was stepped between runs. The latter isn't /too/ bad from a user
perspective, though it's confusing as a developer. We probably will
only end up showing the more recent recording for a given
timestamp anyway. But the former is quite annoying. It means we have
to throw away part of the SourceBuffer that we may want to seek back
(causing UI pauses when that happens) or keep our own spare copy of it
(memory bloat). I'd like to avoid the whole mess.
* Another approach is to use timestamps that are guaranteed to be in
the correct order but that may have gaps. In particular, a timestamp
of (recording_id * max_recording_duration) + time_within_recording.
But again seeking isn't instantaneous. In my experiments, there's a
visible pause between segments that drives me nuts.
* Finally, the approach that led me to this schema change. Use
timestamps that place each segment after the one before, possibly with
an intentional gap between runs (to force a wait where we have an
actual gap). This should make the browser's natural playback behavior
work properly: it never goes to an incorrect place, and it only waits
when/if we want it to. We have to maintain a mapping between its
timestamps and segment ids but that's doable.
This commit is only the schema change; the new data aren't exposed in
the API yet, much less used by a UI.
Note that stream.next_recording_id became stream.cum_recordings. I made
a slight definition change in the process: recording ids for new streams
start at 0 rather than 1. Various tests changed accordingly.
The upgrade process makes a best effort to backfill these new fields,
but of course it doesn't know the total duration or number of runs of
previously deleted rows. That's good enough.
2020-06-09 19:17:32 -04:00
|
|
|
select
|
|
|
|
composite_id,
|
|
|
|
open_id,
|
|
|
|
stream_id,
|
|
|
|
run_offset,
|
|
|
|
flags,
|
|
|
|
sample_file_bytes,
|
|
|
|
start_time_90k,
|
|
|
|
duration_90k,
|
|
|
|
video_samples,
|
|
|
|
video_sync_samples,
|
|
|
|
video_sample_entry_id
|
|
|
|
from
|
|
|
|
old_recording
|
|
|
|
order by composite_id
|
2021-02-17 01:15:54 -05:00
|
|
|
"#,
|
|
|
|
)?;
|
|
|
|
let mut insert = tx.prepare(
|
|
|
|
r#"
|
track cumulative duration and runs
This is useful for a combo scrub bar-based UI (#32) + live view UI (#59)
in a non-obvious way. When constructing a HTML Media Source Extensions
API SourceBuffer, the caller can specify a "mode" of either "segments"
or "sequence":
In "sequence" mode, playback assumes segments are added sequentially.
This is good enough for a live view-only UI (#59) but not for a scrub
bar UI in which you may want to seek backward to a segment you've never
seen before. You will then need to insert a segment out-of-sequence.
Imagine what happens when the user goes forward again until the end of
the segment inserted immediately before it. The user should see the
chronologically next segment or a pause for loading if it's unavailable.
The best approximation of this is to track the mapping of timestamps to
segments and insert a VTTCue with an enter/exit handler that seeks to
the right position. But seeking isn't instantaneous; the user will
likely briefly see first the segment they seeked to before. That's
janky. Additionally, the "canplaythrough" event will behave strangely.
In "segments" mode, playback respects the timestamps we set:
* The obvious choice is to use wall clock timestamps. This is fine if
they're known to be fixed and correct. They're not. The
currently-recording segment may be "unanchored", meaning its start
timestamp is not yet fixed. Older timestamps may overlap if the system
clock was stepped between runs. The latter isn't /too/ bad from a user
perspective, though it's confusing as a developer. We probably will
only end up showing the more recent recording for a given
timestamp anyway. But the former is quite annoying. It means we have
to throw away part of the SourceBuffer that we may want to seek back
(causing UI pauses when that happens) or keep our own spare copy of it
(memory bloat). I'd like to avoid the whole mess.
* Another approach is to use timestamps that are guaranteed to be in
the correct order but that may have gaps. In particular, a timestamp
of (recording_id * max_recording_duration) + time_within_recording.
But again seeking isn't instantaneous. In my experiments, there's a
visible pause between segments that drives me nuts.
* Finally, the approach that led me to this schema change. Use
timestamps that place each segment after the one before, possibly with
an intentional gap between runs (to force a wait where we have an
actual gap). This should make the browser's natural playback behavior
work properly: it never goes to an incorrect place, and it only waits
when/if we want it to. We have to maintain a mapping between its
timestamps and segment ids but that's doable.
This commit is only the schema change; the new data aren't exposed in
the API yet, much less used by a UI.
Note that stream.next_recording_id became stream.cum_recordings. I made
a slight definition change in the process: recording ids for new streams
start at 0 rather than 1. Various tests changed accordingly.
The upgrade process makes a best effort to backfill these new fields,
but of course it doesn't know the total duration or number of runs of
previously deleted rows. That's good enough.
2020-06-09 19:17:32 -04:00
|
|
|
insert into recording (composite_id, open_id, stream_id, run_offset, flags,
|
2020-08-05 00:44:01 -04:00
|
|
|
sample_file_bytes, start_time_90k, prev_media_duration_90k,
|
|
|
|
prev_runs, wall_duration_90k, media_duration_delta_90k,
|
|
|
|
video_samples, video_sync_samples, video_sample_entry_id)
|
track cumulative duration and runs
This is useful for a combo scrub bar-based UI (#32) + live view UI (#59)
in a non-obvious way. When constructing a HTML Media Source Extensions
API SourceBuffer, the caller can specify a "mode" of either "segments"
or "sequence":
In "sequence" mode, playback assumes segments are added sequentially.
This is good enough for a live view-only UI (#59) but not for a scrub
bar UI in which you may want to seek backward to a segment you've never
seen before. You will then need to insert a segment out-of-sequence.
Imagine what happens when the user goes forward again until the end of
the segment inserted immediately before it. The user should see the
chronologically next segment or a pause for loading if it's unavailable.
The best approximation of this is to track the mapping of timestamps to
segments and insert a VTTCue with an enter/exit handler that seeks to
the right position. But seeking isn't instantaneous; the user will
likely briefly see first the segment they seeked to before. That's
janky. Additionally, the "canplaythrough" event will behave strangely.
In "segments" mode, playback respects the timestamps we set:
* The obvious choice is to use wall clock timestamps. This is fine if
they're known to be fixed and correct. They're not. The
currently-recording segment may be "unanchored", meaning its start
timestamp is not yet fixed. Older timestamps may overlap if the system
clock was stepped between runs. The latter isn't /too/ bad from a user
perspective, though it's confusing as a developer. We probably will
only end up showing the more recent recording for a given
timestamp anyway. But the former is quite annoying. It means we have
to throw away part of the SourceBuffer that we may want to seek back
(causing UI pauses when that happens) or keep our own spare copy of it
(memory bloat). I'd like to avoid the whole mess.
* Another approach is to use timestamps that are guaranteed to be in
the correct order but that may have gaps. In particular, a timestamp
of (recording_id * max_recording_duration) + time_within_recording.
But again seeking isn't instantaneous. In my experiments, there's a
visible pause between segments that drives me nuts.
* Finally, the approach that led me to this schema change. Use
timestamps that place each segment after the one before, possibly with
an intentional gap between runs (to force a wait where we have an
actual gap). This should make the browser's natural playback behavior
work properly: it never goes to an incorrect place, and it only waits
when/if we want it to. We have to maintain a mapping between its
timestamps and segment ids but that's doable.
This commit is only the schema change; the new data aren't exposed in
the API yet, much less used by a UI.
Note that stream.next_recording_id became stream.cum_recordings. I made
a slight definition change in the process: recording ids for new streams
start at 0 rather than 1. Various tests changed accordingly.
The upgrade process makes a best effort to backfill these new fields,
but of course it doesn't know the total duration or number of runs of
previously deleted rows. That's good enough.
2020-06-09 19:17:32 -04:00
|
|
|
values (:composite_id, :open_id, :stream_id, :run_offset, :flags,
|
2020-08-05 00:44:01 -04:00
|
|
|
:sample_file_bytes, :start_time_90k, :prev_media_duration_90k,
|
|
|
|
:prev_runs, :wall_duration_90k, 0, :video_samples,
|
|
|
|
:video_sync_samples, :video_sample_entry_id)
|
2021-02-17 01:15:54 -05:00
|
|
|
"#,
|
|
|
|
)?;
|
track cumulative duration and runs
This is useful for a combo scrub bar-based UI (#32) + live view UI (#59)
in a non-obvious way. When constructing a HTML Media Source Extensions
API SourceBuffer, the caller can specify a "mode" of either "segments"
or "sequence":
In "sequence" mode, playback assumes segments are added sequentially.
This is good enough for a live view-only UI (#59) but not for a scrub
bar UI in which you may want to seek backward to a segment you've never
seen before. You will then need to insert a segment out-of-sequence.
Imagine what happens when the user goes forward again until the end of
the segment inserted immediately before it. The user should see the
chronologically next segment or a pause for loading if it's unavailable.
The best approximation of this is to track the mapping of timestamps to
segments and insert a VTTCue with an enter/exit handler that seeks to
the right position. But seeking isn't instantaneous; the user will
likely briefly see first the segment they seeked to before. That's
janky. Additionally, the "canplaythrough" event will behave strangely.
In "segments" mode, playback respects the timestamps we set:
* The obvious choice is to use wall clock timestamps. This is fine if
they're known to be fixed and correct. They're not. The
currently-recording segment may be "unanchored", meaning its start
timestamp is not yet fixed. Older timestamps may overlap if the system
clock was stepped between runs. The latter isn't /too/ bad from a user
perspective, though it's confusing as a developer. We probably will
only end up showing the more recent recording for a given
timestamp anyway. But the former is quite annoying. It means we have
to throw away part of the SourceBuffer that we may want to seek back
(causing UI pauses when that happens) or keep our own spare copy of it
(memory bloat). I'd like to avoid the whole mess.
* Another approach is to use timestamps that are guaranteed to be in
the correct order but that may have gaps. In particular, a timestamp
of (recording_id * max_recording_duration) + time_within_recording.
But again seeking isn't instantaneous. In my experiments, there's a
visible pause between segments that drives me nuts.
* Finally, the approach that led me to this schema change. Use
timestamps that place each segment after the one before, possibly with
an intentional gap between runs (to force a wait where we have an
actual gap). This should make the browser's natural playback behavior
work properly: it never goes to an incorrect place, and it only waits
when/if we want it to. We have to maintain a mapping between its
timestamps and segment ids but that's doable.
This commit is only the schema change; the new data aren't exposed in
the API yet, much less used by a UI.
Note that stream.next_recording_id became stream.cum_recordings. I made
a slight definition change in the process: recording ids for new streams
start at 0 rather than 1. Various tests changed accordingly.
The upgrade process makes a best effort to backfill these new fields,
but of course it doesn't know the total duration or number of runs of
previously deleted rows. That's good enough.
2020-06-09 19:17:32 -04:00
|
|
|
let mut rows = stmt.query(params![])?;
|
|
|
|
while let Some(row) = rows.next()? {
|
|
|
|
let composite_id: i64 = row.get(0)?;
|
|
|
|
let open_id: i32 = row.get(1)?;
|
|
|
|
let stream_id: i32 = row.get(2)?;
|
|
|
|
let run_offset: i32 = row.get(3)?;
|
|
|
|
let flags: i32 = row.get(4)?;
|
|
|
|
let sample_file_bytes: i32 = row.get(5)?;
|
|
|
|
let start_time_90k: i64 = row.get(6)?;
|
2020-08-05 00:44:01 -04:00
|
|
|
let wall_duration_90k: i32 = row.get(7)?;
|
track cumulative duration and runs
This is useful for a combo scrub bar-based UI (#32) + live view UI (#59)
in a non-obvious way. When constructing a HTML Media Source Extensions
API SourceBuffer, the caller can specify a "mode" of either "segments"
or "sequence":
In "sequence" mode, playback assumes segments are added sequentially.
This is good enough for a live view-only UI (#59) but not for a scrub
bar UI in which you may want to seek backward to a segment you've never
seen before. You will then need to insert a segment out-of-sequence.
Imagine what happens when the user goes forward again until the end of
the segment inserted immediately before it. The user should see the
chronologically next segment or a pause for loading if it's unavailable.
The best approximation of this is to track the mapping of timestamps to
segments and insert a VTTCue with an enter/exit handler that seeks to
the right position. But seeking isn't instantaneous; the user will
likely briefly see first the segment they seeked to before. That's
janky. Additionally, the "canplaythrough" event will behave strangely.
In "segments" mode, playback respects the timestamps we set:
* The obvious choice is to use wall clock timestamps. This is fine if
they're known to be fixed and correct. They're not. The
currently-recording segment may be "unanchored", meaning its start
timestamp is not yet fixed. Older timestamps may overlap if the system
clock was stepped between runs. The latter isn't /too/ bad from a user
perspective, though it's confusing as a developer. We probably will
only end up showing the more recent recording for a given
timestamp anyway. But the former is quite annoying. It means we have
to throw away part of the SourceBuffer that we may want to seek back
(causing UI pauses when that happens) or keep our own spare copy of it
(memory bloat). I'd like to avoid the whole mess.
* Another approach is to use timestamps that are guaranteed to be in
the correct order but that may have gaps. In particular, a timestamp
of (recording_id * max_recording_duration) + time_within_recording.
But again seeking isn't instantaneous. In my experiments, there's a
visible pause between segments that drives me nuts.
* Finally, the approach that led me to this schema change. Use
timestamps that place each segment after the one before, possibly with
an intentional gap between runs (to force a wait where we have an
actual gap). This should make the browser's natural playback behavior
work properly: it never goes to an incorrect place, and it only waits
when/if we want it to. We have to maintain a mapping between its
timestamps and segment ids but that's doable.
This commit is only the schema change; the new data aren't exposed in
the API yet, much less used by a UI.
Note that stream.next_recording_id became stream.cum_recordings. I made
a slight definition change in the process: recording ids for new streams
start at 0 rather than 1. Various tests changed accordingly.
The upgrade process makes a best effort to backfill these new fields,
but of course it doesn't know the total duration or number of runs of
previously deleted rows. That's good enough.
2020-06-09 19:17:32 -04:00
|
|
|
let video_samples: i32 = row.get(8)?;
|
|
|
|
let video_sync_samples: i32 = row.get(9)?;
|
|
|
|
let video_sample_entry_id: i32 = row.get(10)?;
|
|
|
|
if cur_stream_id != Some(stream_id) {
|
|
|
|
cum_duration_90k = 0;
|
|
|
|
cum_runs = 0;
|
|
|
|
cur_stream_id = Some(stream_id);
|
|
|
|
}
|
2021-02-17 01:15:54 -05:00
|
|
|
insert
|
|
|
|
.execute_named(named_params! {
|
|
|
|
":composite_id": composite_id,
|
|
|
|
":open_id": open_id,
|
|
|
|
":stream_id": stream_id,
|
|
|
|
":run_offset": run_offset,
|
|
|
|
":flags": flags,
|
|
|
|
":sample_file_bytes": sample_file_bytes,
|
|
|
|
":start_time_90k": start_time_90k,
|
|
|
|
":prev_media_duration_90k": cum_duration_90k,
|
|
|
|
":prev_runs": cum_runs,
|
|
|
|
":wall_duration_90k": wall_duration_90k,
|
|
|
|
":video_samples": video_samples,
|
|
|
|
":video_sync_samples": video_sync_samples,
|
|
|
|
":video_sample_entry_id": video_sample_entry_id,
|
|
|
|
})
|
|
|
|
.with_context(|_| format!("Unable to insert composite_id {}", composite_id))?;
|
2020-08-05 00:44:01 -04:00
|
|
|
cum_duration_90k += i64::from(wall_duration_90k);
|
track cumulative duration and runs
This is useful for a combo scrub bar-based UI (#32) + live view UI (#59)
in a non-obvious way. When constructing a HTML Media Source Extensions
API SourceBuffer, the caller can specify a "mode" of either "segments"
or "sequence":
In "sequence" mode, playback assumes segments are added sequentially.
This is good enough for a live view-only UI (#59) but not for a scrub
bar UI in which you may want to seek backward to a segment you've never
seen before. You will then need to insert a segment out-of-sequence.
Imagine what happens when the user goes forward again until the end of
the segment inserted immediately before it. The user should see the
chronologically next segment or a pause for loading if it's unavailable.
The best approximation of this is to track the mapping of timestamps to
segments and insert a VTTCue with an enter/exit handler that seeks to
the right position. But seeking isn't instantaneous; the user will
likely briefly see first the segment they seeked to before. That's
janky. Additionally, the "canplaythrough" event will behave strangely.
In "segments" mode, playback respects the timestamps we set:
* The obvious choice is to use wall clock timestamps. This is fine if
they're known to be fixed and correct. They're not. The
currently-recording segment may be "unanchored", meaning its start
timestamp is not yet fixed. Older timestamps may overlap if the system
clock was stepped between runs. The latter isn't /too/ bad from a user
perspective, though it's confusing as a developer. We probably will
only end up showing the more recent recording for a given
timestamp anyway. But the former is quite annoying. It means we have
to throw away part of the SourceBuffer that we may want to seek back
(causing UI pauses when that happens) or keep our own spare copy of it
(memory bloat). I'd like to avoid the whole mess.
* Another approach is to use timestamps that are guaranteed to be in
the correct order but that may have gaps. In particular, a timestamp
of (recording_id * max_recording_duration) + time_within_recording.
But again seeking isn't instantaneous. In my experiments, there's a
visible pause between segments that drives me nuts.
* Finally, the approach that led me to this schema change. Use
timestamps that place each segment after the one before, possibly with
an intentional gap between runs (to force a wait where we have an
actual gap). This should make the browser's natural playback behavior
work properly: it never goes to an incorrect place, and it only waits
when/if we want it to. We have to maintain a mapping between its
timestamps and segment ids but that's doable.
This commit is only the schema change; the new data aren't exposed in
the API yet, much less used by a UI.
Note that stream.next_recording_id became stream.cum_recordings. I made
a slight definition change in the process: recording ids for new streams
start at 0 rather than 1. Various tests changed accordingly.
The upgrade process makes a best effort to backfill these new fields,
but of course it doesn't know the total duration or number of runs of
previously deleted rows. That's good enough.
2020-06-09 19:17:32 -04:00
|
|
|
cum_runs += if run_offset == 0 { 1 } else { 0 };
|
|
|
|
}
|
2021-02-17 01:15:54 -05:00
|
|
|
tx.execute_batch(
|
|
|
|
r#"
|
2020-03-20 23:52:30 -04:00
|
|
|
drop index recording_cover;
|
|
|
|
create index recording_cover on recording (
|
|
|
|
stream_id,
|
|
|
|
start_time_90k,
|
|
|
|
open_id,
|
2020-08-05 00:44:01 -04:00
|
|
|
wall_duration_90k,
|
|
|
|
media_duration_delta_90k,
|
2020-03-20 23:52:30 -04:00
|
|
|
video_samples,
|
|
|
|
video_sync_samples,
|
|
|
|
video_sample_entry_id,
|
|
|
|
sample_file_bytes,
|
|
|
|
run_offset,
|
|
|
|
flags
|
|
|
|
);
|
|
|
|
|
|
|
|
alter table recording_integrity rename to old_recording_integrity;
|
|
|
|
create table recording_integrity (
|
|
|
|
composite_id integer primary key references recording (composite_id),
|
|
|
|
local_time_delta_90k integer,
|
|
|
|
local_time_since_open_90k integer,
|
|
|
|
wall_time_delta_90k integer,
|
|
|
|
sample_file_blake3 blob check (length(sample_file_blake3) <= 32)
|
|
|
|
);
|
|
|
|
insert into recording_integrity
|
|
|
|
select
|
|
|
|
composite_id,
|
|
|
|
local_time_delta_90k,
|
|
|
|
local_time_since_open_90k,
|
|
|
|
wall_time_delta_90k,
|
|
|
|
null
|
|
|
|
from
|
|
|
|
old_recording_integrity;
|
|
|
|
|
|
|
|
alter table recording_playback rename to old_recording_playback;
|
|
|
|
create table recording_playback (
|
|
|
|
composite_id integer primary key references recording (composite_id),
|
|
|
|
video_index blob not null check (length(video_index) > 0)
|
|
|
|
);
|
|
|
|
insert into recording_playback select * from old_recording_playback;
|
|
|
|
|
|
|
|
drop table old_recording_playback;
|
|
|
|
drop table old_recording_integrity;
|
|
|
|
drop table old_recording;
|
track cumulative duration and runs
This is useful for a combo scrub bar-based UI (#32) + live view UI (#59)
in a non-obvious way. When constructing a HTML Media Source Extensions
API SourceBuffer, the caller can specify a "mode" of either "segments"
or "sequence":
In "sequence" mode, playback assumes segments are added sequentially.
This is good enough for a live view-only UI (#59) but not for a scrub
bar UI in which you may want to seek backward to a segment you've never
seen before. You will then need to insert a segment out-of-sequence.
Imagine what happens when the user goes forward again until the end of
the segment inserted immediately before it. The user should see the
chronologically next segment or a pause for loading if it's unavailable.
The best approximation of this is to track the mapping of timestamps to
segments and insert a VTTCue with an enter/exit handler that seeks to
the right position. But seeking isn't instantaneous; the user will
likely briefly see first the segment they seeked to before. That's
janky. Additionally, the "canplaythrough" event will behave strangely.
In "segments" mode, playback respects the timestamps we set:
* The obvious choice is to use wall clock timestamps. This is fine if
they're known to be fixed and correct. They're not. The
currently-recording segment may be "unanchored", meaning its start
timestamp is not yet fixed. Older timestamps may overlap if the system
clock was stepped between runs. The latter isn't /too/ bad from a user
perspective, though it's confusing as a developer. We probably will
only end up showing the more recent recording for a given
timestamp anyway. But the former is quite annoying. It means we have
to throw away part of the SourceBuffer that we may want to seek back
(causing UI pauses when that happens) or keep our own spare copy of it
(memory bloat). I'd like to avoid the whole mess.
* Another approach is to use timestamps that are guaranteed to be in
the correct order but that may have gaps. In particular, a timestamp
of (recording_id * max_recording_duration) + time_within_recording.
But again seeking isn't instantaneous. In my experiments, there's a
visible pause between segments that drives me nuts.
* Finally, the approach that led me to this schema change. Use
timestamps that place each segment after the one before, possibly with
an intentional gap between runs (to force a wait where we have an
actual gap). This should make the browser's natural playback behavior
work properly: it never goes to an incorrect place, and it only waits
when/if we want it to. We have to maintain a mapping between its
timestamps and segment ids but that's doable.
This commit is only the schema change; the new data aren't exposed in
the API yet, much less used by a UI.
Note that stream.next_recording_id became stream.cum_recordings. I made
a slight definition change in the process: recording ids for new streams
start at 0 rather than 1. Various tests changed accordingly.
The upgrade process makes a best effort to backfill these new fields,
but of course it doesn't know the total duration or number of runs of
previously deleted rows. That's good enough.
2020-06-09 19:17:32 -04:00
|
|
|
drop table old_stream;
|
2020-03-20 23:52:30 -04:00
|
|
|
drop table old_video_sample_entry;
|
|
|
|
|
|
|
|
update user_session
|
|
|
|
set
|
|
|
|
revocation_reason = 1,
|
|
|
|
revocation_reason_detail = 'Blake2b->Blake3 upgrade'
|
|
|
|
where
|
|
|
|
revocation_reason is null;
|
2021-02-17 01:15:54 -05:00
|
|
|
"#,
|
|
|
|
)?;
|
2020-03-20 00:35:42 -04:00
|
|
|
Ok(())
|
|
|
|
}
|