Skip to content

Latest commit

 

History

History
656 lines (551 loc) · 26.5 KB

MIGRATING.md

File metadata and controls

656 lines (551 loc) · 26.5 KB
Raw

Migrating to Releases with Breaking Changes

Table of Contents

Migrating to v2.1

Unified Links

Over the years, I implemented 4 different versions of the Link entries:

  • Ghost Links (< v1.5)
  • Fat Links (>= v1.5)
  • Thin Links (>= v1.6)
  • Symbolic Links (>= v1.11)

I had mistakenly assumed that TZDB Link entries were somehow less important than the Zone entries. Often Link entries were old spellings of zones which were replaced by new zone names (e.g. "Asia/Calcutta" replaced by "Asia/Kolkata"), or zones using an older naming convention pre-1993 (e.g. "UTC" replaced by "Etc/UTC").

The code in AceTime reflected my assumption of the second-class status of Link entries. Recently however the IANA TZDB project has aggressively merged unrelated Zones (in different countries) into a single zone if they all happen to have the same DST transition rules since 1970. The duplicate zones become Link entries to the "canonical" zone (e.g. "Arctic/Longyearbyen", "Europe/Copenhagen", "Europe/Oslo", "Europe/Stockholm", "Atlantic/Jan_Mayen" are all links to "Europe/Berlin").

It is now more clear the Link entries should be considered first-class entries, just like Zone entries. The v2.1 release implements this change in semantics. All previous implementations of Links are now merged into a single implementation which treats Links equal to Zones, and all the usual operations on Zones are also valid on Links.

  1. The "Thin Link" feature has been removed, along with the Link Manager classes. The code was too complex, and did not provide enough value.
    • LinkManager
    • BasicLinkManager
    • ExtendedLinkManager.
  2. The followLink parameter on various TimeZone and ZoneProcessor methods has been removed.
  3. The zonedb and zonedbx databases no longer contain the link registries:
    • basic::kLinkRegistry
    • extended::kLinkRegistry
  4. The kZoneRegistry is still generated for historical reasons.
    • This registry contains the minimum complete dataset of the IANA timezones.
    • Most applications should use kZoneAndLinkRegistry which contains the Link entries.

Most application code are expected to treat Links and Zones equally. There are only 2 methods which apply only to Link time zones:

  • TimeZone::isLink()
    • Returns true if the current time zone is a Link.
  • TimeZone::printTargetZoneNameTo()
    • Prints the name of the target Zone if the current zone is a link.
    • It prints nothing is isLink() is false.

ZonedExtra

The ZonedExtra class was created to replace 3 ad-hoc query methods on the TimeZone object:

  • Removed: TimeZone::getUtcOffset()
  • Removed: TimeZone::getDeltaOffset()
  • Removed: TimeZone::getAbbrev()

Once the ZonedExtra object has been objected for a particular point in time, the following methods on ZonedExtra are the replacements for the above:

  • ZonedExtra::timeOffset()
  • ZonedExtra::dstOffset()
  • ZonedExtra::abbrev()

The ZonedExtra object will normally be created through 2 factory methods:

  • ZonedExtra::forEpochSeconds(epochSeconds, tz)
  • ZonedExtra::forLocalDateTime(ldt, tz)

The ZonedExtra object provides access to other meta-information about the time zone at that particular time. See the ZonedExtra section in the USER_GUIDE.md for more detailed information about this class.

Migrating to v2.0

High Level

The primary purpose of AceTime v2 is to extend the range of years supported by the library from [2000,2050) to [2000,10000), while remaining compact enough to be useful on resource constrained environments like 8-bit AVR processors. AceTime uses a 32-bit integer to represent the "epoch seconds". This means that it has a range about 136 years. In version v1, the epoch was hardcoded to be 2000-01-01T00:00:00, which allowed the "epoch seconds" to support years from about 1950 to 2050, which got truncated to about 2000 to 2050 for simplicity. However, in the year 2022, the upper limit of 2050 seems too close for comfort to use in embedded devices which could last more than 25 years.

One solution for extending the range of the "epoch seconds" is to use a 64-bit integer instead of a 32-bit integer. This solution is used by many other time zone libraries. However, 64-bit operations are extremely resource intensive on 8-bit processors in terms of both flash memory and CPU cycles, and did not seem appropriate for AceTime which hopes to remain usable on 8-bit processors.

The solution used by AceTime v2 is to keep the "epoch seconds" as a 32-bit integer, but allow the epoch year to be adjustable instead of being hardcoded to the year 2000. Downstream applications can select an epoch year that is appropriate for their use case, allowing AceTime to be valid over roughly a 100-year interval straddling the epoch year. For example, if the epoch year is set to 2100 (so that the epoch is 2100-01-01T00:00:00), AceTime will work for the years 2050 to 2150. The assumption is that a 100-year interval is sufficient for most embedded applications.

With the epoch year being adjustable, it becomes necessary to decouple the various year fields in the TZDB database (implemented by the zonedb and zonedbx packages) from the year fields in the AceTime library code itself. Previously in v1, the year fields were encoded as 8-bit int8_t integers which were interpreted to be offsets from the hardcoded epoch year of 2000. To allow the zone databases to be valid until the year 10000, the internal year fields are changed from an int8_t to a 16-bit int16_t type. The AceTime API hides most of the impact of this change from client applications, so the biggest noticeable change may be the increase in flash size of the zonedb and zonedbx databases which are now 2.5 kiB to 3.5 kiB larger.

The increase in flash size for zonedb and zonedbx seems acceptable for the following reasons: The full impact of the size increase would be felt only if the application incorporated the entire zonedb or zonedbx database to support all 595 time zones in the TZDB database. But most 8-bit processors do not have enough flash memory to use the full database (e.g. 23 kiB for zonedb, 38 kiB for zonedbx), so it is likely that these 8-bit applications would use only a small subset (1-10) of the timezones available in those databases. The flash size increase would be far smaller when using only small number of time zones. On 32-bit processors where the full database would likely be used, they often have far more flash memory (0.5 MiB to 4 MiB on ESP8266 or ESP32), so the increase of 2.5-3.5 kiB of flash memory would be negligible on those processors.

Some backwards incompatible changes were necessary from v1 to v2. These are explained in detail in the next section.

Details

AceTime v2 implements the following major changes and features:

  • the internal year field in various classes (LocalDate, LocalDateTime, OffsetDateTime, ZonedDateTime) changes from int8_t to an int16_t
    • the range increases from [1873,2127] to [1,9999]
    • the various year() methods in these classes were already using int16_t so this internal change should be mostly invisible to client applications
  • the year fields in the zonedb and zonedbx databases also change from int8_t to int16_t
    • the year range increases from [2000,2049] to [2000,9999]
    • decouples the TZ database from the adjustable current epoch year
  • removed constants
    • LocalDate::kEpochYear
      • replacement: Epoch::currentEpochYear() function
      • reason: no longer a constant
    • LocalDate::kSecondsSinceUnixEpoch
      • purpose: number of seconds from 1970 to the AceTime epoch (2000-01-01 in v1, but adjustable in v2)
      • replacement: Epoch::secondsToCurrentEpochFromUnixEpoch64()
      • reasons:
        • int32_t seconds can overflow, so use int64_t
        • epoch year is now adjustable, not a constant
    • LocalDate::kDaysSinceUnixEpoch
      • purpose: number of days from 1970-01-01 to AceTime epoch (2000-01-01 in v1, but adjustable in v2)
      • replacement: Epoch::daysToCurrentEpochFromUnixEpoch()
      • reason: epoch is now adjustable, so must become a function
    • LocalDate::kMinYearTiny
      • replacement: LocalDate::kMinYear
      • reason: 8-bit offset no longer used, replaced by 16-bit integer
    • LocalDate::kMaxYearTiny
      • replacement: LocalDate::kMaxYear
      • reason: 8-bit offset no longer used, replaced by 16-bit integer
    • LocalDate::kInvalidUnixDays
      • replacement: kInvalidEpochDays
      • reason: simplification, both had the same value INT32_MIN
    • LocalDate::kInvalidUnixSeconds
      • replacement: LocalDate::kInvalidUnixSeconds64
      • reason: 32-bit versions of toUnixSeconds() removed
  • removed functions
    • LocalDate::toUnixSeconds()
      • reason: 32-bit Unix seconds will overflow in the year 2038
      • replacement: LocalDate::toUnixSeconds64()
    • LocalDate::forUnixSeconds()
      • reason: 32-bit Unix seconds will overflow in the year 2038
      • replacement: LocalDate::forUnixSeconds64()
    • LocalDate::yearTiny()
      • reason: int8_t year fields replaced by int16_t type
    • LocalDate::forTinyComponents() (undocumented)
      • reason: int8_t year fields replaced by int16_t type
    • OffsetDateTime::toUnixSeconds()
    • OffsetDateTime::forUnixSeconds()
    • OffsetDateTime::yearTiny()
    • ZonedDateTime::toUnixSeconds()
    • ZonedDateTime::forUnixSeconds()
    • ZonedDateTime::yearTiny()
  • new functions
    • Epoch::currentEpochYear(epochYear)
      • purpose: set the current epoch year
    • Epoch::currentEpochYear()
      • purpose: get the current epoch year
    • Epoch::daysToCurrentEpochFromUnixEpoch()
      • purpose: number of days from Unix epoch (1970-01-01) to the current epoch ({yyyy}-01-01) where yyyy is set by currentEpochYear(yyyy)
    • Epoch::daysToCurrentEpochFromConverterEpoch()
      • purpose: number of days from the converter epoch (2000-01-01T00:00:00) to the current epoch ({yyyy}-01-01T00:00:00) where yyyy is set by currentEpochYear(yyyy)
      • comment: should not normally be needed by client applications
    • Epoch::secondsToCurrentEpochFromUnixEpoch64()
      • purpose: number of seconds from the Unix epoch (1970-01-01T00:00:00) to the current epoch ({yyyy}-01-01T00:00:00)
      • comment: useful for converting between AceTime epoch and Unix epoch
    • Epoch::epochValidYearLower()
      • purpose: defines lower limit of valid years (valid_year >= lower) for features related to epochSeconds and timezones
    • Epoch::epochValidYearUpper()
      • purpose: defines upper limit of valid years (valid_year < upper) for features related to epochSeconds and timezones

The epochSeconds that was generated by AceTime v1 using the epoch year of 2000 will be incompatible with AceTime v2 using a different epoch year. Client applications which need read old epochSeconds value using AceTime v2 have a number of options:

  1. Call Epoch::currentEpochYear(2000) at the beginning of the application, so that the v2 epoch year is the same as the v1 epoch year. The disadvantage is that the 32-bit epochSeconds will stop working with this library sometime around the year 2065-2066.
  2. Perform a conversion of the v1 epochSeconds to the v2 epochSeconds by setting Epoch::currentEpochYear(year) first, then calculating the new epochSeconds using newEpochSeconds = oldEpochSeconds - Epoch::daysToCurrentEpochFromConverterEpoch() * 86400.
  3. Do no conversion. Just reset the date and time using the new epoch year. The next time the device is rebooted, the date and time will use the new epoch year instead of the old epoch year.

Background Motivation

Using 32-bit integer field for epochSeconds gives a range of about 136 years. This is the cause of the famous Unix Year 2038 problem which uses a 32-bit signed integer starting from the epoch year of 1970 (1970-01-01 00:00:00 UTC).

When the AceTime project started in 2018, using the year 2000 as the epoch year pushed the theoretical maximum year of epochSeconds to about 2068 which seemed sufficiently far enough away. The epoch year of 2000 also seemed convenient because it is the same value used by the AVR libc in its time.h implementation. The actual upper limit was restricted to 2050 to provide some headroom before calculations would overflow in the year 2068.

Now in the year 2022, the upper limit of 2050 feels too low, since embedded devices could be reasonably expected to keep working for the next 25 years. The updated AceTime v2 is designed to support a 100-year interval from [2000,2100) by default. To prevent the need to change the source code when the range needs to extended even further in the future, the "current epoch year" is made adjustable by the client application.

Migrating to v1.9.0

The ZoneManager hierarchy (containing ManualZoneManager, BasicZoneManager, and ExtendedZoneManager) was refactored from v1.8.0 to v1.9.0.

Configuring the Zone Managers

In v1.8, the ZoneManager was an abstract interface class with 7 pure virtual methods that was the base class of the class hierarchy of all ZoneManager subclasses. This was convenient because the TimeZone related parts of the client application code could be written against the ZoneManager base class and the specific implementation could be configured in a small section of the application code. The problem with such a polymorphic class hierarchy is that the virtual methods consume significant amounts of flash memory, especially on 8-bit AVR processors with limited flash. The examples/MemoryBenchmark program showed that this design consumed an extra 1100-1300 bytes of flash.

In v1.9, several changes were made to reduce the flash memory size:

  1. All virtual methods were removed from the ZoneManager and its subclasses.
  2. The BasicZoneManager and ExtendedZoneManager classes are no longer template classes, making them easier to use (e.g. in the ZoneSorterByName and ZoneSorterByOffsetAndName classes).
  3. The internal BasicZoneProcessorCache and ExtendedZoneProcessorCache member variables were extracted out from the respective ZoneManager classes. These are now expected to be created separately, and passed into the constructors of the BasicZoneManager and ExtendedZoneManager classes.

The migration path is relatively simple. In v1.8, the BasicZoneManager was configured like this:

static const uint8_t CACHE_SIZE = 4;
BasicZoneManager<CACHE_SIZE> zoneManager(
    kZoneRegistrySize,
    kZoneRegistry);

In v1.9, this should be replaced with code that looks like:

static const uint8_t CACHE_SIZE = 4;
BasicZoneProcessorCache<CACHE_SIZE> zoneProcessorCache;
BasicZoneManager zoneManager(
    kZoneRegistrySize,
    kZoneRegistry,
    zoneProcessorCache);

Similarly, in v1.8, the ExtendedZoneManager was configured like following:

static const uint8_t CACHE_SIZE = 4;
ExtendedoneManager<CACHE_SIZE> zoneManager(
    zonedb::kZoneRegistrySize,
    zonedb::kZoneRegistry);

In v1.9, this should be replaced with code that looks like this:

static const uint8_t CACHE_SIZE = 4;
ExtendedoneProcessorCache<CACHE_SIZE> zoneProcessorCache;
ExtendedoneManager zoneManager(
    zonedbx::kZoneRegistrySize,
    zonedbx::kZoneRegistry,
    zoneProcessorCache);

Using the Zone Managers

In v1.8, the ZoneManager was the parent interface class of all polymorphic subclasses. So the client code that needed a specific subclass of ZoneManager could do something like this:

class Controller {
  public:
    Controller(
      ZoneManager* zoneManager,
      ...
    ) :
      mZoneManager(zoneManager),
      ...
    {}

  private:
    ZoneManager* mZoneManager;
};

Any instance of BasicZoneManager<SIZE> or ExtendedZoneManager<SIZE> could be passed into the constructor. This provided some runtime flexibility and code simplicity. However, the runtime flexibility did not seem useful for the vast majority of cases and the simplicity offered by the single parent interface class was paid for by an extra 1100-1300 bytes of flash memory.

In v1.9, the application still has the ability to choose between a BasicZoneManager and an ExtendedZoneManager at compile time. The same Controller constructor should look something like this:

// Define the various TIME_ZONE_TYPE macros in config.h.
#include "config.h"

class Controller {
  public:
    Controller(
    #if TIME_ZONE_TYPE == TIME_ZONE_TYPE_BASIC
      BasicZoneManager* zoneManager,
    #elif TIME_ZONE_TYPE == TIME_ZONE_TYPE_EXTENDED
      ExtendedZoneManager* zoneManager,
    #endif
      ...
    ) :
      mZoneManager(zoneManager),
      ...
    {}

  private:
  #if TIME_ZONE_TYPE == TIME_ZONE_TYPE_BASIC
    BasicZoneManager* mZoneManager;
  #elif TIME_ZONE_TYPE == TIME_ZONE_TYPE_EXTENDED
    ExtendedZoneManager* mZoneManager;
  #endif
};

It is assumed that most applications will hard code either the BasicZoneManager or the ExtendedZoneManager, and will not need this level of configuration.

Link Managers

In v1.8, the LinkManager was an interface class with pure virtual methods:

class LinkManager {
  public:
    static const uint16_t kInvalidZoneId = 0x0;
    virtual uint32_t zoneIdForLinkId(uint32_t linkId) const = 0;
    virtual uint16_t linkRegistrySize() const = 0;
};

But allowing the BasicLinkManager and ExtendedLinkManager to be polymorphic did not seem worth the extra flash usage, similar to the ZoneManager hierarchy.

In v1.9, the pure virtual methods were removed, but the static constant was retained for backwards compatibility:

class LinkManager {
  public:
    static const uint16_t kInvalidZoneId = 0x0;
};

The BasicLinkManager and ExtendedLinkManager should be used directly, instead of through the LinkManager interface. Since Link Managers were introduced only in v1.8, I expect almost no one to be affected by this.

Migrating to v1.8.0

Three breaking changes were made from v1.7.5 to v1.8.0:

  1. The SystemClock and other clock classes were moved to AceTimeClock. This improves the decoupling between the AceTime and AceTimeClock libraries and allows faster development of each library.
  2. The DS3231Clock class was converted into a template class to replace a direct dependency to the I2C <Wire.h> library with an indirect dependency to the AceWire library. This reduces the flash memory consumption between 1300-2500 bytes on AVR processor on applications which use only the AceTime portion of the library, and increases the flexibility of the DS3231Clock class.
  3. Support for thin links was moved out of BasicZoneManager and ExtendedZoneManager into the new BasicLinkManager and ExtendedLinkManager classes. This simplifies the ZoneManagers, and reduces the flash memory consumption of applications which do not use this feature by 200-500 bytes.

The following subsections show how to migrate client application from AceTime v1.7.5 to AceTime v1.8.0.

Migrating to AceTimeClock

For AceTime v1.8.0, the clock classes under the ace_time::clock namespace have been moved to the AceTimeClock library. To help backwards compatibility, the namespace of the clock classes remain in the ace_time::clock namespace.

To migrate your old code, install AceTimeClock using the Arduino Library Manager. (See AceTimeClock Installation for more details). Then update the client code to add the <AceTimeClock.h> header file just after the exiting <AceTime.h> header.

For example, if the original code looks like this:

#include <AceTime.h>
using namespace ace_time;
using namespace ace_time::clock;

Replace that with this:

#include <AceTime.h>
#include <AceTimeClock.h>
using namespace ace_time;
using namespace ace_time::clock;

Migrating the DS3231Clock

For AceTime v1.8.0, the DS3231Clock class was converted into a template class to replace a direct dependency to the <Wire.h> library with an indirect dependency to to the AceWire library. There were 2 primary motivation for the change.

One, simply including the <Wire.h> header file increases the flash memory usage by ~1300 bytes on AVR, even if the Wire object is never used. The DS3231Clock class is the only class in the AceTime library that depended on the <Wire.h>. So any application that pulled in <AceTime.h> for the time zone classes would suffer the increased flash usage of the <Wire.h> library, even if the Wire was never referenced or used in the client application.

Two, the TwoWire class from <Wire.h> is not designed to be used polymorphically (see SoftwareWire#28 for more details). In other words, it cannot be subclassed and cannot be replaced with a different implementation of the I2C protocol. If a 3rd party library contains a direct dependency to the <Wire.h> directly, it is impossible to replace the Wire object with a different I2C implementation (for example, one of the alternative I2C implementations listed in this Overview of Arduino I2C libraries. The <AceWire.h> library solves this problem by using compile-time polymorphism through C++ templates.

Here is the migration process. For all occurrences of the old DS3231Clock class that look like this:

#include <AceTimeClock.h>
using namespace ace_time;
using namespace ace_time::clock;
...
DS3231Clock dsClock;

Replace that with a template instance of the DS3231Clock<T> class, and its I2C interface classes (TwoWire and TwoWireInterface):

#include <AceTimeClock.h>
#include <AceWire.h> // TwoWireInterface
#include <Wire.h> // TwoWire, Wire
using namespace ace_time;
using namespace ace_time::clock;

using WireInterface = ace_wire::TwoWireInterface<TwoWire>;
WireInterface wireInterface(Wire);
DS3231Clock<WireInterface> dsClock(wireInterface);

void setup() {
  ...
  Wire.begin();
  wireInterface.begin();
  dsClock.setup();
  ...
}

The new DS3231Clock<T> class requires more configuration. But in return, we gain more flexibility and potentially a large reduction of flash memory consumption if the pre-installed <Wire.h> is replaced with a different I2C implementation. For example, here is a version of the DS3231Clock object where the SimpleWireInterface software I2C class replaces the hardware TwoWireInterface class, without any changes to the DS3231Clock class:

#include <AceTimeClock.h>
#include <AceWire.h>
using namespace ace_time;
using namespace ace_time::clock;

const uint8_t SCL_PIN = SCL;
const uint8_t SDA_PIN = SDA;
const uint8_t DELAY_MICROS = 4;
using WireInterface = ace_wire::SimpleWireInterface;
WireInterface wireInterface(SDA_PIN, SCL_PIN, DELAY_MICROS);
DS3231Clock<WireInterface> dsClock(wireInterface);

void setup() {
  ...
  wireInterface.begin();
  dsClock.setup();
  ...
}

According to the benchmarks at AceWire/MemoryBenchmark, using SimpleWireInterface instead of the TwoWire class reduces flash consumption by 1500 bytes on an AVR processor. The flash consumption can be reduced by 2000 bytes if the "fast" version SimpleWireFastInterface is used instead.

Migrating to LinkManagers

In v1.7.5, thin links were activated by adding the kLinkRegistrySize and kLinkRegistry parameters to the constructor of BasicZoneManager and ExtendedZoneManager, like this:

BasicZoneManager zoneManager(
    zonedb::kZoneRegistrySize, zonedb::kZoneRegistry,
    zonedb::kLinkRegistrySize, zonedb::kLinkRegistry);

ExtendedZoneManager zoneManager(
    zonedbx::kZoneRegistrySize, zonedbx::kZoneRegistry,
    zonedbx::kLinkRegistrySize, zonedbx::kLinkRegistry);

This caused BasicZoneManager::createForZoneId() and ExtendedZoneManager::createForZoneId() to perform an additional lookup in the kLinkRegistry if a zoneId was not found in the kZoneRegistry.

In v1.8.0, that fall-back functionality has been moved to the LinkManagers, and the ZoneManager constructors no longer accept the link registry parameters:

BasicLinkManager linkManager(
    zonedb::kLinkRegistrySize, zonedb::kLinkRegistry);

ExtendedLinkManager linkManager(
    zonedbx::kLinkRegistrySize, zonedbx::kLinkRegistry);

The client application is now responsible for activating the link registry, and performing the fallback lookup:

TimeZone findTimeZone(uint32_t zoneId) {
  TimeZone tz = zoneManager.createForZoneId(zoneId);
  if (tz.isError()) {
    // Search the link registry.
    zoneId = linkManager.zoneIdForLinkId(zoneId);
    if (zoneId != LinkManager::kInvalidZoneId) {
      tz = zoneManager.createForZoneId(zoneId);
    }
  }
  return tz;
}

This change allows the ZoneManagers to provide a consistent API for some upcoming features, and prevents unnecessary flash consumption (200-500 bytes) if the client application does not use the thin link feature.